PSA: how to use GPGPU

• Step 0: slurm does not know about the GPUs!
• Step 1: check GPUs

​

​

• Step 2: set environment variable to ensure consistent indexing

​

​

• Step 3 (option A): set requested GPU to a free one

​

• Step 3 (option B): choose requested GPU automatically

​

• Step 4: start computation
• Step 5: check that everything went according to plan

In general: do not leave jupyter notebooks idle while occupying GPU vram

Wishlist (to learn/discuss in the future) pls keep at top

• Foundation model
• Multi-modal
• Common practice in e.g. plotting, data analysis, packages
• Maybe ask mathematicians and statisticians to help us learn when needed
• Code design workshop
• KG: write a CD3 GPT
• How does moving mesh etc work?

Done

• Summarize Astro/Cosmo experiments people involved in IPMU (?) CD3 Project Expo May 10, 2024 Website
• Parallel Computing / HPC (during CD3 opening symposium)
• Flow-based models - Joaquin
• Energy based learning model - Thomas (abondoned)
• Parallel computing on idark (Connor’s talk) - https://github.com/cbottrell/HPC_IPMU
• Letter on stop researching AI for 6 months link1, link2
• Instruction to use globus (Wenke, Hideki)
• A3Net summer school materials: https://github.com/IPMUCD3/a3net_2024
• Transformer – Sherry

​

Pause for summer, continue in Fall 2024. ML lecture (webpage, videos)

HAIC Slides

Friday Jan 31

HAC preparation

Dec 18th 2024, Ben, Jia

Plan

• Book 5F or 4F seminar room from Jan (Jia to do)
• Pair up projects (7 projects!)
• Ben create slide deck for HAC (link)
• List of tools (Ben, Jia)
• ChatGPT Plus subscription (20 people x $20/mo x 12 months = $4800/yr)
• Literature review / learning the background and past works (typeset.io, elicit.com, you.com, xyzt.ai, jenni.ai, researchrabbit.ai)
• Design the project and code up the main analysis tools (chatGPT, copilot, claude, gemini…)
• Summarize and write up the paper (chatGPT, claude, )
• Workflow
• First session (~mid Jan): Ben describe rules (human can’t use AI); project benchmark
• Everyone describe the project, work through the block with their companion, by the end, present steps
• Second session: literature review, a test run get familiar with all the tools; by the end share experience
• Usual hack should still happen for people to report and discuss their projects
• Timeline
• Announce the project assignment and room change, first meeting, and format (Ben, Jan 6, to announce Jan 24 or 31st first meeting)

CD3 Hack Friday

Dec 13th 2024, 1:30-16:30pm 3F seminar room

What do you want to explore this afternoon?

• Katya: make morphology comparison plots for active/non-active galaxies - not there yet, but had a good start, will continue after weekend
• Joaquin: Make plots for SBI validation (still reading papers and documentation). (TARP-pablo lemos, Lampe).
• Leander: work on FAIRS-Japan white paper, write on introduction for cluster mass ML paper (w/ Bonny)
• Boris: Finally start trying to implement ML based classification on spiral features.
• Linda: debug my simulation code that is killed by MPI abort for no reason
• Kevin: Develop ML technique to identify spectral class from Euclid grism sims
• Sherry: Implement my transformer code on gpgpu
• Joao: Learn very basics of scikit-learn

​

Katya

CD3 Hack Friday

Nov 29th 2024, 1:30-16:30pm 3F seminar room

What do you want to explore this afternoon?

• Jia: prepare a 10 course syllabus for the GUC summer school.
• Cesar: Do nice plots for paper / Think about HAC
• Katya: re-check fitting results after fixing the bug in my code
• Jess: talk to leander/jia, add new mark functions to my code class
• Jingjing: talk to Minh, Kota
• Sherry: Try to make my CNN work longer than short length.
• Joaquin: Finish poster for workshop next week.
• Leander: talk w/ Jess & Jia, work on ACT experiment, think about Princeton postdoc.

Jia’s course outline

Intro: Through this course, we will explore the history of our universe and its components. We will also practice using the tools scientists employ to study the cosmos. Emphasis will be placed on developing essential skills for scientific research—problem-solving, critical thinking, and computational analysis.

10 topics: big bang and a brief history, dark matter, dark energy, galaxies, stars, black holes, planets, scientific measurements, scientific programming, and statistics.

• Made plots :)

Leander

Can’t show much from ACT…

​

CD3 Hack Friday

Oct 11th 2024, 1:30-16:30pm 3F seminar room

What do you want to explore this afternoon?

• Cesar: Working with A. Muñoz; we are reviewing neutrino detector data / simulation agreement (essentially debugging).
• Katya: plot BH mass vs Galaxy mass and other main relations
• Jess: think about covariances,
• ​
• Akira: ask Leander how to use halomodel PDF code
• Chenze: Working on the field-level inference code (lightcone stuff)
• Jia: Learn QFT - ML connection.
• Baptiste: optimize matrix multiplication / inversion

​

CD3 Hack Friday

Oct 4th 2024, 1:30-16:30pm 3F seminar room

What do you want to explore this afternoon?

• Linda: Preparing presentation for COSMO’24. 30min slides+ 10min (Wed)
• Broad audience. Lots of people!
• Feedback welcomed!
• Overview of numerical studies.
• Finished a couple slides, (finding some animation; finalizing flow chart.)
• Baptiste: check component separation step/map->spectra step for SO pipeline. Coding.
• Fixed the Megatop bug. Merged disagreements.
• Akira: Make duty slide for supervisors.
• Jess: 5min Talk on Wednesday for France/Japan Science community.
• Broad audience.
• Big people are there.
• Make presentation on cosmology in 5min. Feedback welcomed if before 3pm. Weak lensing animation needed.
• Rec letter??
• Jia: (sorry I’m at Kavli visiting event), but I’ll try to learn about current understanding of double descent curve in ML, and possible ways QFT can help （sherry：gotcha👍
• Joaquin: Organizing CD3 hackathon; talking about planning.
• Sherry: Explore Planck raw TOD(time-ordered-data) and cleaned data; see how humans did it first to narrow down the architecture choice for the pipeline.
• Try flagging first.

CD3 Hack Friday

Sept 27 2024, 1:30-16:30pm 3F seminar room

What do you want to explore this afternoon?

• Katya: Continue with CIGALE, talk to John. Realize how to get FIRST catalogue.
• Joaquin: Jump into a new project, looking on the data I have access (simulations, nodes, emulators, SBI). Talk to Eiichiro.
• Linda: continue debugging my clustering dark energy simulation code. Read Joaquin’s paper.
• Cesar: Mainly writing paper, might do quick debug test with Junjie
• Akira: beautify the final plots of lensing SSC. figure out why the kappa map from tiled boxes looks weird
• Ben: Make slides describing new Differentiable FOF methods

Katya

CD3 Hack Friday

Sept 20 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Katya: Figure out how to simulate fluxes in CIGALE
• Akira: mask equator of Healpix map and re-measure the stats
• Joaquin: Preparing test (plot) for non-Gaussianity of the Minkowski functional HSC data.
• Linda: continue debugging my clustering dark energy simulation code

NG MFs - Joaquin

Found some high S2N for different fields (cutting the edges)

More NG features when combining fields

Akira - Stats of equator-masked Healpix map

Q: Sub box from bigbox and tile it as for small box

​

It will have same replication effect.

Katya

CD3 Hack Friday

Sept 13 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jia: study the general use of AI in particle physics
• Linda: Implementing a different Riemann solver in my clustering dark energy simulation code
• Leander: KMI workshop, pension, prepare for LtU Workshop (inferring cosmology from galaxy survey)
• Katya: keep running CIGALE (SED fitting code) to find more appropriate qso hosts mass estimations

Topics of interest for cosmology x ML

— implicit-likelihood inference / simulation-based inference / likelihood-free inference

— emulators for summary statistics

— field-based surrogate models (e.g., baryon painting)

— differentiable simulations

​

Anything missing?

Jia: spent whole afternoon admiring lectures from Lukas Heinrich, TUM

CD3 Hack Friday

July 26 2024, 1:30-5pm 3F seminar room

Gender Diversity in STEM in Japan: Challenges & Solutions with Insights from Researchers

Shawn Mobley

2024-07-26

Japan is awesome

But it struggles with gender equality

Many universities: undergraduate student body 20% female

For context

Especially in STEM

Following statistics: Sasada & Bannai, 2021 – mathematics

Estimates: High school around 20%, Undergraduate around 15%

Compared to select OECD nations

Share of females keeps diminishing

Left with 2 separate but related problems: entrance and retention

​

Many possible causes and many possible solutions

Many projects underway

Ikkatai et al., 2021

Ceci et al., 2009

Ceci et al., 2009

Wanted to hear from researchers

• Goals:
• Contextualize/model the pipeline that causes gender imbalance in Japan
• Compare results from literature in Japan and abroad with interviews
• Assess certain proposed solutions and recommend which ones to use
• 20 interviews: 14 female, 6 male
• Several major universities across Japan
• Questions about childhood, education, career, experience with discrimination, and perspectives on society
• Around 1 hr (but as short as 30 mins and as long as 1.5 hrs!)
• Recorded and transcribed: text = primary source

Some initial findings / observations

• Profs are extreme right tail of “thick skin”: not necessarily representative
• Many went to all-girls schools, had parents in academic/science, did not perceive much/any discrimination, had not given solutions much thought
• Discrimination examples: very direct (“women shouldn’t be here”), moderate (difficulty with hiring, passive aggressiveness), very indirect (8pm meetings)
• Everyone recognized need for change and had hope in younger generation
• Most agreed quotas were necessary, but they weren’t happy about it
• Despite low birth rate, motherhood and family remain very culturally important
• South Korea is complicated to consider: increased representation but anti-feminism, gender divisions, …
• Curious to learn more about Japanese education system, especially exams
• Boys and girls tend to do well on different aspects of courses and exams
• Also curious about demographic trends (marriage, having children, …)
• Tenure comes at worst possible time for women, whose impact dips but then recovers later
• Interesting but complicated to look at older literature from US/Europe

Some considerations

• Evaluating data at both the mean and the right tail (men & women have different distributions)
• Unclear which metrics best predict (let alone cause) success in STEM fields
• Many seemingly problematic factors in Japan are also present elsewhere
• Need to challenge assumptions, think very logically, and avoid stereotypes

Next steps

• Thoroughly analyze interview transcripts
• Keep reading literature on Japan and other countries
• Write a report on findings (approx. 5000 words)
• Hopefully submit to new GUNDAM journal!
• Continue onto second phase: testing large language models for gender bias

Shoutouts

• Profs. Jia Liu & Hiromi Yokoyama
• Prof. Jun’ichi Yokoyama & IPMU
• The incredible student researchers at IPMU
• The 20 professors I interviewed
• Profs. Saeko Hayashi, Yukari Ito & women’s lunch members
• Profs. Takao Hensch, Gavin Whitelaw & Harvard RIJS

My travels around Japan

Thank you!

Astro Visualization

​

Bonny Y. Wang

https://github.com/Cyber-Astro-Telescope

Interactive Examples

Web Interactive Example

• https://bonnywang.github.io/AstroData/
• https://www.tng-project.org/explore/3d/
• https://3d-asteroids.space/asteroids

​

Particle Data

Density Field Data

https://github.com/Cyber-Astro-Telescope/Tutorials

https://www.blender.org/

​

Basic Setup + Cheatsheet

• Edit -> Preferences -> Emulate Numpad
• Two finger / hold the middle wheel to rotate
• Shift + Two finger / hold the middle wheel to move
• Zoom in + zoom out
• ` to view selected and view camera view
• ctrl + alt + 0 to align camera to the current view
• G to Grab, S to scale and R to rotate
• Shift + A to Add new things
• X to delete an object

BlenderExamples/IPMU_Workshop/MergerGas_4.npz

BlenderExamples/DensityFeildVis.py

​

• Change root and file
• Root = “{Your Local Path to Tutorial }/BlenderExamples/IPMU_Workshop/”
• File = MergerGas_4.npz
• return data["arr_0"] -> return data[“gas”]

Work with Density Field Animation

Use

BlenderExamples/IPMU_Workshop/gas_{0-100}.vdb

Shift + A -> Volume -> Import OpenVDB

• Shading
• Try different rendering method
• Add camera trajectory and render the video

Render on Cluster

wget https://mirror.freedif.org/blender/release/Blender4.2/blender-4.2.0-linux-x64.tar.xz

tar -xf blender-4.2.0-linux-x64.tar.xz

# Render the first frame

./blender-4.2.0-linux-x64/blender -b file.blend -f 1

# Render the animaiton

./blender-4.2.0-linux-x64/blender -b file.blend -a

Other Useful Things

• https://github.com/Cyber-Astro-Telescope/Tutorials/blob/main/BlenderExamples/PositionAnimation.py
• https://www.youtube.com/watch?v=fnbhYxj8D_s

​

CD3 Hack Friday

July 12th 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jia: read a review on primordial magnetic fields
• Chenze: prepare slides for the intro to cosmic density reconstruction
• Leander: continue writing paper about debaryonification with optimal transport
• Jingjing: read BAO reconstruction paper (have to leave at 4:30 to pick up baby)

Leander: debaryonification

Float Precisions

Always 10^(original magnitude - 8)

CD3 Hack Friday

July 5th 2024, 1:30-5pm 3F seminar room

PSA: how to use GPGPU

• Step 0: slurm does not know about the GPUs!
• Step 1: check GPUs

​

​

• Step 2: set environment variable to ensure consistent indexing

​

​

• Step 3 (option A): set requested GPU to a free one

​

• Step 3 (option B): choose requested GPU automatically

​

• Step 4: start computation
• Step 5: check that everything went according to plan

In general: do not leave jupyter notebooks idle while occupying GPU vram

What do you want to explore this afternoon? (5/7)

• Jia: understand the Milky Way magnetic field
• Leander: write some introduction about de-baryonification via optimal transportr, make some plots maybe
• Zhaoxuan: write some introduction for my thesis on starburst galaxies at cosmic noon
• Akira: Read Agora paper,write shell rotation code, try GRAYTRIX
• Joaquin: Talk to Akira about Flask, Jia, and finish draft paper about Minkowski funcs.
• Katya: check the robustness of Galight fit to slightly different PSFs
• Omar: Work on gpgpu to train diffusion model.
• Bonny: make some visualizations? (show this in 2 weeks)
• ​

Jia: https://arxiv.org/abs/2304.12350

Leander: wrote some introduction and made a nice figure

Observations of a grand-design spiral starburst galaxy at z=1.463

CD3 Hack Friday

June 28th 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jia: contemplate how to teach undergraduate neutrino oscillation, numerical simulations, and AI/ML
• Ben: MCLMC (Fancy HMC (Fancy MCMC)) combined with JaxPM (sample ICs and/or cosmological parameters)
• Joaquin: Work in project with Valentin. Chat with Jia and Leander.
• Akira: generate lognormal map (Flask), measure HOS of GSN noisy map
• Jess: fix emulator for new mark functions
• Leander: implement more physical metric, talk w/ Cooper & Jia
• Kanmi: CNN (now studying ML)
• Bonny: look for improvements can be done for the cluster mass prediction
• Omar: Implement diffusion model

Akira - Fibonacci Grid (link)

How to chop as many flat patch from full-sky

Some NFW integrals - Leander

Jess :(

Jia, playing with JaxPM: https://github.com/DifferentiableUniverseInitiative/JaxPM/blob/main/notebooks/Introduction.ipynb

What are these 2 fields?? ^

Why use galaxy dynamics to measure cluster mass

• Not completely virialized
• Traditionally known as the M-σ relation.
• High-mass tail of the halo mass function (HMF)
• Cluster abundance constrain cosmological models
• Cosmological parameters
• Neutrino Mass

M_Sigma relationship

• Add layernorm and dropout for preventing overfitting

CD3 Hack Friday

“Project Brainstorming” Informal Discussion

June 21st 2024, 1:30-5pm 3F seminar room

Name - Project description

Please introduce the motivation and general picture of your data-analysis/simulation/ML project. Please also emphasize the part new people can join, or other groups can collaborate with.

Each project should take 5-8 min and a couple of slides..

Subaru HSC – weak lensing

• All members at IPMU have the full data rights
• Subaru HSC: all data acquisition (~1100 sq. deg) done. Currently the data processing ongoing. The next cosmology papers are for the full dataset (final)
• Shape catalog and photoz info will be ready by the end of Aug (TBD): Xiangchong (CMU⇒Brookhaven) & Tianqing (Pitts)
• Validation and null tests will be done by Oct (TBD): Hironao, Sunao, Surhud, … you are welcome!
• Physics analyses of the shape catalog will be after the validation tests. Key projects would be led by members who have expertise (but based on the collaboration policy). You are welcome for new analyses, e.g. using higher-order moments or SBI
• Challenges – simulations & new methods (e.g. bispectrum, SBI, …)
• N-body sims (AbacusSummit, 1000 sims of 4096^3 & 1(Gpc/h)^3 (Tomoki Ishiyama @ Chiba)
• Full-sky (or light-cone) mocks: Takahashi sims, new sims by Ken Osato (Chiba U.)
• Hydrosims – galaxy-halo connection, or other validation tests
• Developments of “estimators” – e.g., bispectrum estimator (Taniguchi & Terawaki)
• IPMU contacts: Masahiro Takada

Subaru PFS – BAO, redshift-space distortion, full-shape …

• All members at IPMU have the full data rights
• Will start PFS survey in Feb, 2025 (will cover the HSC-Wide of ~1100 sq. deg)
• Challenges in obs side – you are very welcome!
• Catalog making: HSC targets ⇒ fiber assignments ⇒ successful redshifts. Details of all these steps need to be fixed (Shintaro, Jingjing, MT, ….)
• Observation strategy: exposure time, tiling of different fields, …. TBD (Shintaro, Jingjing, …)
• Characterization of the selection function: random catalog generation (Shintaro, Jingjing, …)
• Estimators of power spectrum – this would be okay (many methods are in market)
• Challenges in theory/analyses – you are welcome!
• PFS cosmology tracers: emission line galaxies ([OII]), not LRGs. ELGs are not well known
• ELGs tend to reside in less massive halos (down to ~10^11Msun). The mocks require high-resolution simulations
• Analytical method — EFTofLSS
• Simulations based method – ???
• IPMU contacts: Masahiro Takada, Jingjing Shi

Simulations that MT is aware of

Field Level Inference with Subaru PFS and FRBs (KG)

Project I: FRBs and spectroscopic surveys

Use dispersion measure DM=∫ n_e dl from FRBs to probe cosmic baryon distribution

Example of field-level inference: build bespoke IGM+CGM models for each FRB sightline

TBD: (i) NN (or other ML)-based models trained directly on hydro sims (ii) Developing new likelihood framework

Simba simulations (Credit: D. Sorini)

Project II: Subaru PFS GE Survey

PFS Galaxy Evolution Survey will:

• Obtain very high space density of spec-zs over footprints equivalent of modern cosmological hydro boxes + 3D IGM tomography at z~2.5
• Multi-wavelength data available
• Huge redshift baseline (0.7<z<6)
• Density reconstructions/constrained realizations on galaxy data (MCLMC: Chenze Dong) and IGM tomography (Ben Horowitz)

Long-term goal: direct connections between galaxy evolution models and cosmology. E.g. Ben H is developing a differentiable cosmological hydro code

HSC Y3/Y6 weak lensing simulation-based-inference / higher order statistics

• Motivation: extract maximum cosmological information from HSC Y3/Y6 data
• Current progress: HSC Y1 non-Gaussisan stats analysis:
• first cosmological constraints using PDF and scattering transform.
• Improved constraints by up to 30-40% (peaks, minima, PDF, Scattering transform)
• Upcoming: Minkowski Functionals (Joaquin), Marked power spectrum (Jess)
• In planning and welcome collaboration: Y3/Y6 analysis
• Urgent need: simulations
• who, where (Shall we collaborate with LSST DESC and use their sims?)
• Flatsky approximation still ok?
• When will Y6 catalogue done? (xiangchong: needs 2 months after NAOJ sends co-adds)
• Simulation based inferences? (then we need a few x 1000 sims, instead of 100s)
• Sample w, mnu?
• IPMU contacts: Leander Thiele, Jia Liu, Joaquin Armijo, Jess Cowell

HSC Y1 weak lensing higher-order stats

Thiele+2023

Marques+2024

Grandon+2024

Cheng+2024

HalfDome Cosmological Simulations https://halfdomesims.github.io/

• Motivation: neutrino mass, dark energy, PNG
• joint analysis of CMB x LSS surveys (PFS, DESI, LSST, Euclid, Simons Observatory, LiteBIRD..etc.)
• future extension to x-ray, 21cm, FRB, and Ly-alpha
• Current progress
• Nbody runs done: 11 (+1 fNL) simulations of 3.5Gpc, 5x10¹² Msun, z=[0,4]
• Galaxy catalogue in progress (targeting LRGs and ELGs)
• CMB foregrounds: t/kSZ (Linda), radio (Junjie), CIB, lensing (Akira)
• 3 neutrino implementations
• Understanding the super-sample covariance contribution to lensing (Akira Tokiwa)
• Coordinate transform between any 2 arbitrary systems (e.g. Eulerian to spherical) (Cooper Jacobus)
• In planning and welcome new contributions:
• X-ray (cluster cosmology with e.g. eROSITA?)
• 21cm (intensity mapping for Pk, potential in PNG from HI x CMB lensing/CIB?)
• FRB (understand baryonic feedback? need to talk with KG..)
• Ly-alpha (maybe resolution is too far off.. Need ML with superresolution imagine)
• Simulation based inference: collaboration with SIMBIG (2000 fastPM runs)
• IPMU contacts: Jia Liu, Linda Blot

HalfDome Cosmological Simulations https://halfdomesims.github.io/

* Inputs from SO, CMB-S4, LSST, DESI, PFS, SPHEREx, Roman collaborators

HalfDome Cosmological Simulations https://halfdomesims.github.io/

LSST Y1 simulations for higher-order stats (Joaquin, Jia)

Water Cherenkov (WC) Detector Principle

CD3 Project Brainstorming - Patrick de Perio 113

PhysicsOpenLab.org

Neutrino Interactions

CD3 Project Brainstorming - Patrick de Perio 114

T2K (~HK)

Accelerator

Neutrino Flux

Hyper-K Range of Sensitivity

Increasing energy

IBD

CCQE

CC RES

CC DIS

O(1)-O(10) MeV

O(100) MeV

O(1) GeV

O(10) GeV

WC Event Topologies

CD3 Project Brainstorming - Patrick de Perio 115

CC RES

v̅_l

e⁺

p

n

W^-

IBD

ν_l

l^-

n

p

W⁺

CCQE

ν_l

l^-

p

p

W

π⁺

CC DIS

LowE

HighE

10 MeV

e^-

1 GeV

e^-

A rich problem for �event reconstruction:

• Particle classification
• Kinematics regression
• Multi-particle segmentation

Analysis Pipeline

“Event reconstruction”: A method that encodes our physics knowledge for quickly inferring the original particle properties from detector data

CD3 Project Brainstorming - Patrick de Perio 116

Tuning/training�MC

Detector Physics Simulator

Calibration data

Physics�MC

Physics�data

Reconstructed MC

Reconstructed data

Physics analysis

Reconstruction

Reconstruction algorithm tuning

training

ML model

Event Reconstruction

Detector Physics Modeling

Current Paradigm of Detector Physics Modeling

• Limitations
• Lack of “end-to-end” optimization
• Some models are not even optimizable (e.g. look-up tables)
• Same physics, two separate software (i.e. simulation & calibration)
• Goals toward “detector systematics @ <1% level”
• Automation + fast compute that can scale for Hyper-K
• Accurate model optimized directly to minimize data/MC disagreement

CD3 Project Brainstorming - Patrick de Perio 117

Detector Physics Simulator

Detector Output

(ADC)

Simulation Input

(true dE/dX)

Reconstructed

(Calibrated dE/dX)

…

Geometry�

Cherenkov physics

​

Water properties (light scattering, absorption)

​

PMT and wall reflectivity

​

Residual magnetic fields

​

PMT+electronics response

Automation of Detector Physics Model Tuning

Research Proposal: differentiable detector physics simulator (DDSim)

• “End-to-end”: gradient-based optimization using control (calibration) dataset
• Interpretable: analytical physics models for well understood physics
• Flexible: neural representation to incorporate complex features in real data
• Fast: utilization of modern computing accelerators (e.g. GPUs)

CD3 Project Brainstorming - Patrick de Perio 118

Detector Physics Simulator

Key Elements:

1. Differentiability
2. Neural network (NN)

Detector Output

(ADC)

Simulation Input

(true dE/dX)

Reconstructed

(Calibrated dE/dX)

Software Frameworks and Logistics

• Event Reconstruction
• WatChMaL (Water Cherenkov Machine Learning)
• ResNet (CNN), PointNet, GraphNets
• More architecture exploration welcome
• Differentiable Detector Simulator
• CIDeR-ML (Calibration and Inference of Detector Response with Machine Learning)
• SIREN, Taichi Lang
• Collaborators welcome
• Unofficial simulations are openly available
• Application to Super-K data possible, would require either:
• Joining the collaboration (~2 weeks of shift / year), or
• Special permission (including e.g. one-time authorship to resulting paper)
• Application to Water Cherenkov Test Experiment (WCTE) @ CERN easier
• Charge particle beam data coming by end of 2024

CD3 Project Brainstorming - Patrick de Perio 119

CD3 Hack Friday

June 14th 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jia: read literature on magnetic field in and around halos
• Leander: run more hmc
• Katya: remake simulated galaxies images
• Akira: make nicer plots for covariance matrix
• Jess: update intro of paper draft (Optimal mark pk) with new references
• Bonny: build dataset for cluster mass measurement
• Ben: Dealing with annoying Fourier space transformations to power spectra.

​

​

Jia: summary of https://ui.adsabs.harvard.edu/abs/2023A%26A...670L..23H/abstract

HMC runs with different choices of likelihood parameter. Need to wait and then stare.

CD3 Hack Friday

May 31th 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Katya: Fixing things after upgrading packages Galight uses. Aggregate catalogue of fitted sources and make some plots.
• Joaquin: test inference with contaminated data vectors. Help Valentin with his presentation.
• Linda: test new initial conditions code for my dark energy simulations SZ maps for HalfDome
• Jess: have meeting at 4pm, debug contours from cobaya for MPk, junior seminar admin emails
• Akira: read weak lensing analysis

​

Linda: tSZ and kSZ for HalfDome

​

Linda: tSZ and kSZ for HalfDome

CD3 Hack Friday

May 24th 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Linda: converting simulation files to apply SZ painting code
• Akira: read papers about Healpix window function and how it affect the statistics
• Jess: Try to debug cobaya contours (why do they hit the priors)
• Katya: re-writing my wrap classes to Galight in order to solve huge memory consumption

Akira -

Kappa map generated from tiled box has higher value in zs = 3.0

CD3 Hack Friday

May 17th 2024, 1:30-5pm 3F seminar room

Diffusion Models (CS198-126 Lecture 12) - Cesar

• Recap on generative models
• Diffusion
• Some examples / Bonus

A quick recap on other generative models (GANs)

A quick recap on other generative models (VAEs)

A quick recap on other generative models (Flows)

Diffusion

Summarized comparison

Fwd & Bwd in Difussion

Forward Process

Backward Process

Some details on how to actually do it

About the Architecture and Loss

Training and Usage

Some example

Stable diffusion

Bonus

https://stablediffusion.fr/prompts#google_vignette

https://huggingface.co/spaces/multimodalart/stable-video-diffusion

What do you want to explore this afternoon?

• Cesar: No hack today, admin work
• Jia: prepare a talk + discussion on the future of simulations for next week’s conference (cosmo stats)
• Joaquin: make progress on MachineLearning – WeakLensing presentation for next week.
• Linda: read BigFile format in C++ code to paint Sunyaev-Zeldovich signal on halos from HalfDome simulations.
• Omar: Implement diffusion model.
• Junjie: Debug differentiable network for neutrino reconstruction.
• Akira: project maps from spherical to flat sky (coordinate transformation issue)

CD3 Hack Friday

Apr 19th 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Linda: writing grant proposal - I wrote something
• Leander: should be writing grant proposal - installed word
• Tilman: read up on optimal transport, stochastic interpolants - read something and opened notebook
• Giovanni: produce full sky maps of convergence and tSZ from light-cones - produced one octant
• Angela: extrapolate density profile with NFW outside r200 - see plot
• Akira: try to fix the problem on dask-distributed when running wlen code - I tried!

Angela:

GAN

https://arxiv.org/abs/1406.2661

GAN mode collapse

VAE – quick and easy stochasticity

https://arxiv.org/abs/1912.02762

basic autoregressive flow

https://github.com/bayesiains/nflows

https://github.com/vincentstimper/normalizing-flows

https://arxiv.org/abs/1806.07366

Continuous normalizing flow

topological restrictions

Out-of-distribution fail

CIFAR10

SVHN

Normalizing flow issues

CD3 Hack Friday

Apr 5th 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Leander: GTN, make density boxes around IllustrisTNG halos
• Jia: understand link between initial condition and late time displacement field (arxiv)
• Katya: hope to reach the lustre, if not succeed then poke ML a bit more
• Andrés: Rework sample selection script for pionless neutrino interactions in WAGASCI-BabyMIND (T2K near detector)
• Jess: Get cobaya (cosmology sampling code) working (debugging pipeline)
• Boris: Join the hands on session. Go for a run to enjoy sakura blooming in the park.🌸�(to think about https://arxiv.org/pdf/2403.18458.pdf)
• Akira: check if the diffusion model trained properly, plot kappa (projected density map) map
• Cesar: Working a bit on semantic segmentation exercises, also working on differentiable water cherenkov simulation pipeline. Talk to Junjie about Neutrino Earth Tomography with Hyper Kamiokande.
• Tilman: work on Dark Energy Science Collaboration baryon challenge
• Junjie: discussion (Jia on CMB simulations), discussion (Cesar on neutrino tomography) - Yes
• To discuss (Jia): special hack day (all projects); ML current trend review once a month?

Jia: did not learn about Lagrangian perturbation theory, but learned about the Commander pipeline in CMB

https://arxiv.org/pdf/2201.03509.pdf

CD3 Hack Friday

Mar 29 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Leander: reply to referee about voids paper.
• Joaquin: Explore GowerStreet simulations CosmoGrid (weak lensing).
• Akira: deal with MPI code for wlen
• Jess: harass Joaquin about cobaya (likelihood sampler code)
• Camila: Start summarizing some results in a draft.
• Katya: write class to process galight fit results

CosmoGrid sims

Full sky simulations (lightcone) of matter distribution

Gnomonic projection

Smoothing

CD3 Hack Friday

Mar 22 2024, 1:30-5pm 3F seminar room

Hands on Deep Learning

Link

Camila: Meeting

Katya: poking the ML notebooks

CD3 Hack Friday

Mar 15 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Katya: tell about pre-training and data augmentation, write class for spatial SED fitting for Boris
• Jia: learn about velocity field reconstruction from density fields
• Chenze: discuss with Metin on the detail of density field reconstruction
• Leander: figure out lensing maps w/ Akira; void catalog cross matching (optimal transport?)

​

​

​

CD3 Hack Friday

Mar 8 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Boris: Listen to the ML lecture and keep working on prepping galaxy data
• Katya: Enjoy ML lecture. Finish SED fitting
• Angela: Generate some fake data for the electron density profile likelihood
• Jess: Learn from Jia, prepare intro for junior talks, look into good presentation skills
• Linda: ML lecture and fighting with polychord (I lost the fight)
• Andrés: Enjoy Jia’s great lecture. Question quality of my fits.
• Akira: check if the autoencoder is properly working
• Wenke: ML lecture talk (batch norm too) and the refine fitting for IMBH candidate specs.
• Ryo: Make plots comparing HSC-Y3 cosmic shear and Planck.
• Leander: listen to ML lecture (batch norm?), find voids in galaxy mocks

​

​

​

Leander: got VIDE to work on the galaxy catalogs

example QUIJOTE 1Gpc/h box

CD3 Hack Friday

Mar 1 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Katya: listen to ML introduction, run CIGALE on known data to check for biases
• Linda: listen to ML lecture
• Boris: Make some (any) progress on the data prep for ML galaxy structure proj.
• Cesar: Extend Earth Tomography analysis framework, discuss about it with Junjie. Talk about different project with Andrés (WGBM det syst).
• Akira: Figure out how to fix the problem while running MPIsort
• Chenze: Test MCLMC code (git Link) / learn more on JAX
• Andrés: Developing detector systematics for one of the T2K near detectors (WAGASCI-BabyMIND), fight code and listen to Boris’ incredible ML lecture

​

Cesar - I

• Earth Tomography Framework validation looks GOOD! Green vs Blue

https://arxiv.org/pdf/2311.05105.pdf

Cesar - II

• Chat GPT shows how to train on the MNIST – 97% accuracy: 1min code 2s train.

I want to do some classification of the MNIST dataset using keras, provide code for it

• Input to chatGPT3.5

Open projects (anyone welcome to join/contribute) �Please list one project per bullet point for anyone to join, and point to the description page (day of the hack you presented)

​

• Past projects - Abhiram Kidambi (see May 19, 2023 slides), 3 MathxML projects:
• (1) Singularities of Feynman Integrals;
• (2) Primes, quantum chaos, and beyond;
• (3) Classifying solutions to string theory (the Landscape)

CD3 Hack Friday

Feb 16 2024, 1:30-5pm 3F seminar room

Schedule for ML topics (1-slide, 5mins max)

ToDo

• CD3 slack
• Invitation link: https://join.slack.com/t/ipmucd3/shared_invite/zt-2d5wlplwg-13mxx3zvMpRzZxMTkm37_g
• Suggestions for ML/Stats courses to be followed at CD3 hack.
• Adding descriptions of CD3-based projects + Encourage people to start CD3 related projects
• Research which courses are available in ML or statistics
• Next week: break legos

What do you want to explore this afternoon?

• Katya: Prepare list of possible courses Done, finish report draft almost done
• Linda: look for courses with Katya
• Boris: CD3 project idea about predicting stellar structure of galaxies using Deep Neural Network. Also prepare the announcement email with Cesar about new format of Hack fridays.
• Jia: create a cd3 slack, write a proposal w/ john, create example page for project planning
• Cesar: Talk to Boris Katya & Linda. Do a cross-check on final plot result.
• Akira: test healpix generation tasks (super resolution imagine)
• Joaquin: Ask Katya about how to send a merge mail for conference attendees. Done.. Provide Camila more data for HSC project.
• Camila: work on HSC project and try solving/understanding some bias appearing in ML prediction.

​

ML courses:

Stanford machine learning (2022, 20 x 1.2 hours)

Machine learning for physics (2020, 11 x 1.5 hours)

Stanford computer vision (2017, 16 x 1 hour)

Harvard Intro to AI (2024, 6 x 1 hour)

AI-statistics and Astro-ML for beginners (2023, 10 x 1.5hours) (Good for more basic statistics mixed with ML and specific for Astro)

Introduction to Deep Learning (2018-2023, 63 video, ~40mins, a bit random, MIT)

Applied Machine Learning (Cornell Tech CS 5787, Fall 2020)

Berkeley Computer Vision (2022, 22 x ~45 min) (Linda’s favorite in terms of breath)

Fun Deep Reinforcement Learning hands-on course

Akira

The strategy for iterative generation on healpix map

used binary & quaternary number

​

Camila

How to come up with a project (Jia’s version)

• 2 types of projects
• Something you already know well, extension of existing project
• New field (focus of this slide)
• Forming project ideas on a relatively new field:
• Get a general overview: journal search key word and find (1) several review papers, (2) ~10 most cited paper on the topic, (3) ~10 recent papers
• For each paper, read the caveat and future direction part -> a list of unsolved problems
• Pick a few unsolved problems that you can potentially have an interesting attack, check if people have worked out a solution (more journal search)
• Solidify your project idea
• For 2-3 rough problems, come up with a rough plan to solve them (data/simulation/theory/code)
• Talk with experts (tea time!) to vet your ideas (have people done that? Is there obvious bottleneck?)
• Do an order of magnitude estimate of the outcome you can expect (e.g. during hack friday)
• If still interesting, make a more detailed, month by month plan!

CD3 Hack Friday

Feb 9 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Boris: Prepare JWST image based measurements of spatial information of galaxies condensed into a few values. This is to be used for training a neural network to predict stellar structures.
• Katya: Talk to Matt and Bruna, start workshop report draft → started report draft
• Linda: finish research statement for application
• Akira: check different wlen script, trythat. Previous one has memory issue.

CD3 Hack Friday

Feb 2 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Katya: finish my wrapping class to work with Galight, my goal is to make code look nicer and more convenient to make changes - Done!
• Jingjing: Discussion with Natali, Yin; Maybe finish the plots for IA in Parity-violating initial condition simulations
• Cesar: Discuss inputs & framework organization with Junjie for new project (Earth tomography with atmospheric neutrino oscillations with HK). Read related papers. Diff project: Discuss with Andres about WGBM detector systematics.
• Camila: plug HSC Non-Gaussian statistics into my NN code. - Almost Done!
• Jess: get an emulator working for marked power spectra using HSC sims based on Joaquins code, (almost done)
• Akira: try to run weak lensing analysis on idark (it killed my process…).
• /var/spool/pbs/mom_priv/jobs/219440.idark.SC: line 22: 4999 Killed python -m scripts.wlen $DST $SRC $zs --zlmin 0.0 --zlmax 4.0 --zstep=0.1 --nside=8192
• Junjie: try fixing the bug in the neural network for detector optics in the CIDeR-ML project. Discuss with Cesar about the new project. -> bug half-fixed, discussed with Cesar.

Katya (with help of Chat GPT 3.5)

+config file

Junjie

old

new

target

prediction

CD3 Hack Friday

Jan 19 2024, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jia: understand effect of void shape on the void halo mass function (i.e. adrian’s paper)
• Andrew: more c++ solver stuff
• Joaquin: Run marked power spectrum on Quijote-fR suite – Code tested in 1 simulation (need to vectorize now).
• Jingjing: python version of my IA power spectrum code
• Katya: make 2FA for Github, print AI workshop poster, improve decomposition of galaxies with companions
• Linda: work with Adrian on clustering dark energy implementation in FastPM
• Jess: talk to Jia + Joaquin and run Marked Cl on idark of fiducial HSC sims
• Hideki: Prep answer to referee’s question on a paper of component separation with ML�(https://github.com/IPMUCD3/Globus)
• Zhaoxuan: Run grizli to do data processing for a HST PRISM dataset
• Akira: read sample kappa map and measure statistics.
• Adrian: Speak to Kazuyuki about field-level inference. Speak to Linda about CDE in FastPM. Speak to Akira about lensing SSC. Speak to Masahiro about field-level SSC (4pm). Make sure Jia reads my paper.
• Boris: Compiling COSMOS-Web data to implement a (very vague) ML project about finding concentrated star-formation in galaxies. (Ask Zhaoxuan for help)

Possible projects for 2024

• Jia: Faraday Rotation Map from Planck
• Jingjing: which 3pt/bispectrum configuration gives us the most information for 2pt+3pt constraints
• Linda: Clustering dark energy project: forecasts on the speed of sound of dark energy perturbations, analytic computations of splashback radius and void size function, analysis of simulations (later on)

HST G141 FOV

Zoom-in

HST F160W

Direction of dispersion

H alpha

Proposed collaborative project:

Faraday Rotation Map from Planck (Contact: Jia, Toshiya)

source

https://en.wikipedia.org/wiki/Faraday_effect

• Earth B field ~ 1 Gauss
• Galactic B field ~ micro Gauss
• Intergalactic B field << nano Gauss

​

Big question:

• What is the origin of the intergalactic magnetic field (IGMF)
• We can build a Faraday Rotation map from Planck to investigate!

Boris: Finding galaxy cores hidden from view

Similar to near-IR JWST image

HST UV+Optical image contour

(with Dust emission in color)

missing flux

Mass peak

Science: How do galaxies form bulges?

Aim: Predict which galaxies have hidden mass cores without JWST near-IR images

CD3 Hack Friday

Dec 15, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jia: discuss and learn about Intergalactic Magnetic Field and define a student project
• Jingjing: discuss with Junsup via zoom; join the small concert
• Katya: try mail merge; try filling the masked pixels script maskfill from paper (inpainting is similar) – hasn’t even opened maskfill script; sent notification though
• Boris: Make the finishing touches to my paper on the study of galaxy clumps using JWST. Think of an ML galaxy morphology project (again).
• Hideki: Testing super resolution code (LR DM -> HR DM same as Akira)
• Camila: Start studying Normalizing flow.

UV near-IR

CD3 Hack Friday

Dec 1, 2023, 1:30-5pm 3F seminar room

A brief introduction to Symbolic Regression

​

Crash course by César

What is symbolic regression?

We are all more than used to regression analysis, a.k.a doing ‘fits’ to data.

However, what happens if we don’t know what function we should use to fit the data?

​

→ Symbolic regression is a technique that allow us to search in the space of mathematical expressions for the model that best fits the data both in terms of accuracy and simplicity.

​

→ Analytic expressions well-reproducing data can provide direct intuition on the behaviour of a system without human bias and alleviate limitations of discrete data (e.g. analytical differentiation/integration or sampling).

​

​

How does it work?

Symbolic regression is extremely powerful… however: it is an NP-hard problem, namely, a very difficult problem computationally → a solution to your problem is not guaranteed.

To deal with the immense number of possibilities the ideas is to work with expression trees. The building block in symbolic regression are (typically) variables (x,y,z…) binary operators (“+”, “-”, “/”, “*”), unary operators (sin, cos, exp…), and constants.

The basic idea is to use a genetic algorithm:

​

Start with some random expression trees consisting of few variables, operators and constants.

​

Check fitness of each expression (cost function).

​

Select best expressions → generate new expression from them.

​

Next iteration (until convergence criteria is met).

​

Genetic algorithms

Genetic algorithms are inspired in biology. Each iteration defines a ‘generation’ consisting of a pool of ‘candidate solutions’. In every iteration expressions are individually tuned or combined, sorted by fitness, replace older generation.

https://arxiv.org/pdf/2305.01582.pdf

How to define ‘Fitness’?

Imagine we define some complexity for each operator. Then one can calculate the complexity of any given expression:�X: complexity 1 X*X: complexity 2 cos(X)*cos(X)+x^3: complexity 5… etc!

In general, simplest is better. But also we can calculate the distance from the expression to the data, e.g. MSE. We have two variables… so which one do we choose?

Look at the pareto front!

http://fenn.freeshell.org/Science.pdf

How to get started with symbolic regression?

Let’s assume you want to try it for yourself, see how it works and explore if it fits your problem. How hard it is to get started?

​

I recommend getting started with PySR, very nice implementation and user friendly. Implemented by Miles Cranmer (who will visit IPMU in January).

​

→ In less than 1 hack session you can do some tests online using this fantastic demo in google collab, install PySr and start doing your own tests (I did it…).

Bonus: you can use PySR to learn analytic expressions from trained neural networks!

If you are interested checkout this seminar by M. Cranmer on YouTube.

​

What do you want to explore this afternoon?

• Katya: Discuss plots of AGN types with John
• César: Do some numerical tests about what computational speed up can be achieved using factorization in event re-weighting.
• Akira: Prepare Fugaku Machine learning slides
• Wenke: Try the host decomposition consistency on poorer SNR spectra
• Hideki (a.k.a Dr Tanimura): Try Optuna (ML hyperparameter optimization software)
• Chenze: figure out how to obtain 1D power spectrum of Lya forest from CLAMATO data (especially the uncertainty estimation)
• Andrew: continue learning c++ by building solver
• Jess: refresh self on calculating Pk numerically (nbodysims) + comment code and find 2pi factor
• Camila: figure out (or improve) a model independent fitting of P(k) to determine the turning point (where it has a peak).
• Angela: Write cosmosis3 likelihood module of the cluster electron density profile code
• Emile : prepare QUPosium poster (comments welcome!)
• Boris: Trying to understand some data results (galaxy stuff as always).

Katya:

Histograms are showing the overlap between my classification of objects and catalogue types.

Boris

Akira

Jess :(

Camila:

César:

Complexity 4:

Complexity 10:

Claytor et al. 2021

CNN

w/ log-laplacian Adam optimizer

CD3 Hack Friday

Nov 24, 2023, 1:30-5pm 3F seminar room

CD3 x DLX

Special LEGO Hackathon

​

Nov 24 13:30, Kavli IPMU 3F

Hi There!

CD3 x DLX lego workshop

© 2023 DLX Design Lab

We are an international team of designers collaborating with researchers, industry, and the public �to bridge the gap between academia and society.

LOCATIONS

MEMBERS

PORTFOLIO

• Komaba II Campus, S-building
• Kashiwa II Campus, Cooperation hub

25+ people:

• 8 Researchers
• 6 Students
• 4 Admin staff
• 2 Visiting Researchers
• Interns

• 40+ Projects portfolio
• 6 active developments

© 2023 DLX Design Lab

What do we do

EVERYTHING

WE DO IS IN

COLLABORATION

Defining product propositions from early stage technologies

Developing speculative scenarios based on fundamental research

Communicating research to ease understanding and spread public awareness

PRODUCT

VISION

DISSEMINATION

OMNI

​

Collaborator: 藤井研究室（生産技術研究所）

​

Low cost and open-source platform to gather

and analyse critical data from the seas

​

OMNI Microplastics

Citizen science project about engaging more people in the problem of ocean microplastics.

© 2023 DLX Design Lab

TALKING WITH NEURONS

​

Collaborator: 池内研究室（生産技術研究所）

​

An interactive installation that enables people to remotely connect with live neural tissues.

"Talking with Neurons" invites audiences to conduct a reciprocal conversation through sending voice inputs and receiving audio-visual representations of neuronal activity by in vitro neurons that respond to their vocal cues.

© 2023 DLX Design Lab

Every collaboration starts with a workshop

“The best way to have a good idea is �to have a lot of ideas”

​

Linus Pauling

What is a workshop?

A group interactive activity that generates ideas through active participation in structured exercises.

Workshop etiquette

Keep it visual!

Generate creative ideas

Workshop etiquette

Keep it visual! physical!

Generate creative ideas

Not a Lego competition!

Today’s plan

TODAY’S PLAN

Find a team mate

Build!

Tea break

20’

60’

30’

Wrap up

Show/Tell

30’

30’

PRESENTATION TIME

CD3 Hack Friday

Nov 17, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jia: read about intergalactic magnetic field (https://arxiv.org/abs/astro-ph/0408314)
• Cesar: play with symbolic regression using PySr. Do some tests and study project viability for approximating neutrino cross section model generators.
• Katya: deal with emails and invitation letters for the January workshop.
• Andrew: start c++ n-body solver
• Hideki: Read TNG-cluster paper (https://arxiv.org/abs/2311.06338).
• Joaquin: Learn the COBAYA.
• Jess: try to use idark for the first time and find + open HSC data, (need to ask Joaquin for directory maybe? (https://arxiv.org/pdf/2204.05435.pdf) .
• Camila: to decide which ML to use in the case of small parameter space.
• Matt: Set up and run the next batch of Galight decomposition.
• Dongdong: Use MCMC to constrain the sigma(r) with improved experiment parameters of AliCPT; Modify the draft of this forecast paper; Apply a account for the IPMU server.
• Boris: trying to compile ideas about galaxy bulges affecting star-formation in galaxies (for a paper I am writing).

Actual data

Simulations

GAS

STARS

Boris

Cesar

COBAYA - Joaquin

CD3 Hack Friday

Nov 10, 2023, 1:30-5pm 3F seminar room

Normalizing flows for probabilistic modeling and inference

Deep mind paper: 1912.02762

Crash course by Joaquin

Probability density transformation

From base distribution to target distribution

p(u)

p(x)

Transformations: T(u)

How do we define the transformations T if we don’t know how p_u and p_x are related?

​

We can try to model p(u) applying transformations to x

Invertible transforms

The Jacobian

If T is invertible and differentiable then we can compose them:

Flows, inference, and maximum likelihood estimation

p(x) → p(x;θ) fitting a target distribution function p_x(x)

​

KL (Kullback-Leibler) divergence measures the expectation of p(x;θ) explaining the target distribution. For a sample {x_0,x_1,...,x_n} from p_x(x). We can estimate the expectation:

​

​

​

Where the fit can be estimated using maximum likelihood.

​

Params. from transformation T^-1

Params. from base distribution

Constructing flows: Finite composition

First type of flows: Autoregressive flow

With 𝛕 invertible:

This kind of transform. Has a triangular matrix Jacobian

Implementing flows

Transformers

Conic combination can include (non-linear) invertible activation functions: sigmoid, tanh, leaky ReLU.

​

Implementing flows

Conditioners

Coupling layers:

Coupling layers try to approximate the choosing of h and c creating a function F (can be a neural network).

Probabilistic modeling

From gaussian base(s):

What do you want to explore this afternoon?

• Jingjing: prepare my code for calculating halo shapes/angular momentum in Quijote simulation - working progress
• Andrew: refactor halo eigenvalue summing code - is slow
• Chenze: learn some basics on typst (https://typst.app/docs/guides/guide-for-latex-users/ ) & continue working on my (IGM) paper draft - Time to write the discussion section but have no idea how
• Katya: check the point source or host detection in my fits of QSO’s - planned to use maximal likelihood test, but realized I haven’t not reduced chi square. Read a lot about BIC comparison of models, but had no time to implement… Hence Failed…
• Boris: Revise what I can do with the proposed ML based spiral galaxy project .
• Joaquin: Fix my NF neural network. Reply to emails. – Mails done! Looking into previous commits to restore my code!
• Hideki: Train SDSS spectra and hopefully show the prediction
• Jia: understand message passing in Graph neural network (GNN)
• Dongdong:submit tasks on server & revise my paper draft on a new parameterization for constraining narrowband band GWs
• Jess: check stability of my covariance matrix results by removing cross-correlations in covariance matrix (how to neatly code covariance matrices?)
• Angela: Do some Nside=512 lognormal mock (flask) tests for DES Y6 and check the agreement with Maglim
• Camila: will think about her first hack project
• Zhenya: CTA (Cherenkov Telescope Array) gamma ray telescope simulation to work out the observation depending on telescope pointing

Jia: message passing in GNN

Typst

https://typst.app/docs/guides/guide-for-latex-users/

CD3 Hack Friday

Oct 27, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Jingjing: test my code and prepare the slides for the RESCUE meeting
• Akira:examine the result of filter-boost method in sphericalSRdiffusion
• Hideki: modify the GNN paper to include Jia’s major comment.
• Jia: slides for RESCUE meeting
• Andrew: fix plummer sphere code
• Katya: filter out fits with point source not in the center of the cutout

​

Katya: automatic center detected offset. Solved by pinning the initial model center to the center of the cutout

What kind of tool do you use when debugging python code?

import pdb; pdb.set_trace()

%debug

CD3 Hack Friday

Oct 20, 2023, 1:30-5pm 3F seminar room

Hideki

Akira: Vision Transformer

From here

Token: 1d array & position

Scaled dot-product attention

Read the 0-th head (about class)

From here

Not trainable

trainable

Available code: here

What do you want to explore this afternoon?

• Jia: understand probability estimation with normalizing flow (bonus: affine coupling): paper
• Cesar: Understanding/Playing with fitQun pattern recognition algorithm in preparation for ML project. -> fitQun not really implemented in python :(
• Junjie: write a section in T2K technote; might also start reading a paper https://arxiv.org/pdf/2304.01266.pdf
• Baptiste: make plot and run cases for polarization angle requirements in SO (not sure if I can share the plot at the end, probably confidential :P )
• Joaquin: Check compression algorithm for MFs statistic – Reading COBAYA Documentation, as almost all the methods come from there. Give a look to moped.
• Hideki: Run Super Resolution 3D from Akira.
• Andrew: fix sph harmonics code

Akira: Super-resolution diffusion on 3d dark matter density

Baptiste: make plot from test run for polarization angle requirements

(no ticks on the axis for you all, no spoilers :p )

To do:

• Run for all frequency channels (any of you need idark this weekend?)
• Run some more extreme case, see if blue method (r + beta angle) breaks at some point

CD3 Hack Friday

Oct 13, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

​

• Andrew: more ultra light dark matter (ULDM) solver stuff
• Joaquin: Explore codes that apply smoothing in Fourier space and use it in my codes (using jax or Julia).
• Cesar: Read some papers about cherenkov scintillation light separation.
• Akira: Make jupyter notebook of SRDiffusion on 3D grid data
• Jess: work out why cant ssh from desktop on IPMU wifi :(, then compute covariances for 100 COLA sims
• Hideki: Writing a paper of GNN

CD3 Hack Friday

Oct 6, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

​

• Jingjing: do mcmc sampling (emcee) to get the fitting parameters of my luminous red galaxy measurement
• Andrew: more uldm solver stuff, make a solver than runs on the gpu
• Joaquin: Talk to Jess. Solve issues on my LSST codes. – chat with Jess. Checked. LSST pipeline is outdated and I need to add new information provided by new mock maps.
• Wenke: Finish Introduction part of my paper
• Hideki: Structure identification (Node, filament, wall, void) with UNet
• Jess: fix my s8 and om error improvement ratios + draft of my first year report
• Ryo: implement Cell window convolution with Wigner 3-j

Akira: Vision Transformer

From here

Token: 1d array & position

Scaled dot-product attention

Read the 0-th head (about class)

From here

Not trainable

trainable

Available code: here

Structure Identification with reconstructed DM and Galaxies (Hideki)

64^3 grids for 128^3 Mpc/h box

Structure Identification with reconstructed DM and Galaxies (Hideki)

32^3 grids for 128^3 Mpc/h box

CD3 Hack Friday

Sept 29, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Cesar: I work on finalizing ongoing neutral current single positive pion production cross section measurement.
• Jingjing: write fitting code for my measurement of IA of HSC LRGs.
• Joaquin: Update code for Rubin obs./LSST collaboration (HOS topical team).-- I will update my task because the cluster is down :(. Testing emulator for minkowski functionals.
• Andrew: write scalar field solver using eigenvalue decomposition
• Wenke: Tune my AGN-host decomposition code on a small SDSS sample and submit to iDark for DR16 quasar catalog. -> The code is done but have some trouble when submit.
• Hideki: I will write a paper with GNN.

Testing emulator for Minkowski functionals – Joaquin

First unit test:

Replicate f(parameter) for all the parameters used in training.

I have a 10% error in the largest bin, which comes from the individual realizations, Probably driven by shot noise. I need to check how much is my error budget for individual sims.

​

Second unit test.

Replicate f(parameter) for parameter not in the training set, using the omitting one method.

Host Decomposition QA image - Wenke

Eigenvalue decomp sfdm sims - Andrew

Initial conditions code, testing a few different algorithms

<- halos density slices

CD3 Hack Friday

Sept 22, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Katya: Discuss fit results with John, start writing review on AGN feedbacks in MaNGA - done!
• Akira: Try ViT for super resolution & prepare notebook for SRDiffusion
• Joaquin: Mastering YT for data visualization. – Done!
• Jingjing: download the TNG-50 data for MW-like subhalos https://www.tng-project.org/data/docs/api/#getting_started
• Hideki: Reproduce Albert results to prepare for a paper
• Jia: understand density estimation from normalizing flow
• Wenke: Write a notebook introducing who to read out SDSS spectra files (380GB, 4 hr download). Also a handbook to build personal globus endpoint; Instruction to SDSS eboss data usage; SDSS DR17 BOSS location on iDark: /lustre/work/wenke/SDSS/;
• Zhaoxuan: fix my pixel to pixel SED fitting code (https://pixedfit.readthedocs.io/en/latest/index.html). Read the tutorial of MIRI MRS

To do in general for data storage: a dedicated folder on idark for downloaded data for general use by IPMU

Data visualization using YT

With baryons

CD3 Hack Friday

“A safe place to experiment and fail”

Sept 15, 2023, 1:30-5pm 3F seminar room

What do you want to explore this afternoon?

• Cesar: Try to plot a Galaxy
• Jingjing:discussion with Jia/Joaquin/Hideki.
• Boris: Finish at least one sub-section of my paper.Talk about an image processing + ML project
• Jia: investigate Transformer x specs.
• Xuheng: Prepare the slides for the conference.
• Joaquin: Help Jia to learn Transformer and application to specs. Training Diffusion model.
• Hideki: discussion with Jia/Joaquin/Jinjing + try diffusion model if there is time.

Spec x transformer: literature review

• ML papers (Jia):
• Transformer x stellar specs
• Attention-based emulator for specs
• Radio galaxies
• Contrastive learning
• Traditional methods for spec analysis (Joaquin):
• Standard SED fitting procedures.
• Interpreting SEDs of galaxies (BEAGLE).
• CIGALE: Code Investigating GALaxy Emission.
• The Art of Measuring Physical Parameters in Galaxies. Compiles all the above.

Spec-surveys

1. https://vizier.cds.unistra.fr/
2. https://www.astro.ljmu.ac.uk/~ikb/research/galaxy-redshift-surveys.html

CD3 Hack Friday

July 28, 2023

Your goal this afternoon

• Jingjing: write a few emails; reply a few emails (Cesar, I meant SOM, Self-organizing maps)
• Cesar: Understanding splines non-linearities.
• Rhea: look through some old analysis code, label images
• Andrés: Continue working on my poster for the (International Neutrino Summer School) INSS and maybe finish (probably)
• Manolya: complete spectra
• Antoine : write report

​

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

CD3 Hack Friday

July 21, 2023

Your goal this afternoon

• Jingjing: prepare netflow run code and apply to a slice of HSC data
• Sunao: revise HSC weak lensing cosmology paper
• Cesar: Wait for jobs to finish, read papers.
• Hideki: mapping galaxy distribution in redshift space to DM distribution in real space with UNet (wait for the training to finish).
• Antoine : Write more of the report, adapt old code for new data analysis
• Thomas: make new extended source mask.
• Akira: find out the way to run FastPM on Fugaku
• Rhea: submit reimbursement forms and look for images with light injectors out of set of ~1600
• Angela: Optimize the time consumption and precision (N_side) of the galaxy overdensity-template value binning function.
• Cooper: Look at AGN pictures on Galight, make color images, remove point source
• Zhaoxuan: test my MPI script for CASA

​

​

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Mapping galaxy distribution to DM density field

Galaxy density field

DM density field

Reconstructed dm density field without RSD

True

Prediction

Corrcoef = 0.65

Reconstructed dm density field with RSD

True

Prediction

Corrcoef = 0.57

Light injectors (these are highlighted)

Rhea G.

CD3 Hack Friday

July 14, 2023

Your goal this afternoon

• Katya: read carefully paper about simulation of galaxy images. And hopefully start fitting the simulated images with Galight.
• Julien : (working remotely from Kyoto ^^) : need to correct beaming of my CMB maps (there is a resolution problem in my maps, and that biases my cosmological likelihood, I need to correct the beaming to cancel the bias)
• Thomas: generate, assess and verify updated DM mass density distributions.
• Jia: read james’ draft on effects of neutrino mass & dark energy on cosmic web
• Jingjing: construct the catalog for intrinsic alignment measurement of HSC LRGs
• Akira: make the fastpm (n-body simulation code) binary on Fugaku and run sample
• Linda: understand problem with initial conditions of my clustering dark energy simulations
• Boris: Make plots/images to convince John that my results on clumps in galaxies are worth writing a paper on.
• James: While jia is reading over the draft, I’m working on a personal project involving higher order statistics and odd-parity universe… and play chess while my code is running

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Boris: Clumps in star-forming galaxies using JWST

CD3 Hack Friday

July 7, 2023

Your goal this afternoon

• Katya: compare different supersampling settings within galight.
• Thomas: process flux space data for new Fermi LAT IRF(instrument response function); produce new galactic diffuse emission template.
• Jia: for cosmology-dep baryonic feedback, check Astrid halo catalogue
• Andrés: Speak with Junjie about Ptheta (T2K+SK analysis framework) and maybe do some slides while I wait for some jobs to finish
• Jingjing: talk with Xiangchong about the HSC s19a shape catalog
• Akira: see the output of intermediate layers, make perceptual loss
• Hideki: Make a density field from MultiDark snapshot (heavy…)
• Show different gridding schemes (hist, Cloud in Cell (CIC), Triangular Shaped Cloud (TSC))
• Cesar: Preparations for T2K CM.
• Xuheng: Learning HST Grism data reduction
• Sunao: HSC Image reduction - microlensing (M31, 6 yr observation)
• Junjie: prepare talks for T2K CM -> making slides, brainstorming new ideas, talked to Andres about some technical stuff

​

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Jia:

​

Tried to match halos between hydro and dark matter only runs

​

Based on 6D phase (position + velocity)

​

Try pos+velocity fit, kinda ok,

with vw=0.01 for now ->

CD3 Hack Friday

June 30, 2023

Your goal this afternoon

• Julien : write internship report
• Thomas: finish two sections of phenomenological quantum gravity review paper. → (Yona)
• can check this https://indico.cern.ch/event/1279592/ out!
• Jingjing: read random catalog building papers (for PFS data)
• Jia: organize CAMELS halo mass function and understand the effect of baryons on them (also exist on idark jingjing’s directory)
• Rhea: Write some Geant4 simulation scripts (sorry listening to another talk right now)
• Katya: talk to John, plot cute fit results comparison plots -> not cute yet…
• Akira: fix the bug in diffusion model, meeting with Jia, write response to referee
• Andrés: Finish script that merges 2 covariance matrices, needed as input new software I’m using
• Antoine : Make more plots for IWCD analysis with my files ; fix c++ problem with function argument
• Yona: finish the report (add biblio, correct some typos), read sociologist/philosopher/biologist feedback + proofread Albert’s report if time. Also say goodbye :)
• Hideki: To find simulation to mock SDSS galaxies to run UNet to map the relation between galaxy-DM relation.

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Jia, change in halo mass function in CAMELS simulation, due to baryonic feedback

Katya: made comparison plots for two supersampling, need to do more plots for luminosity bins

CD3 Hack Friday

Jun 23, 2023

Your goal this afternoon

• Julien : I made an almond cake! Beware if you’re allergic to eggs, butter, almond (duh) and if you have diabetes… Also, trying to see why I have a bias in my cosmological likelihood
• Junjie: make slides, enjoy the cake - progressing on slides; nice cake thx
• Katya: further experiments with galight fitting, now on cluster!
• Andrés: Make python script to submit jobs in Lyon cluster (slurm)
• Yona: launch the ELUCID cross-correlation at z=0 + get some cleaner figures for the report :)
• Jia: kSZ project order of magnitude estimation; talk james/peter
• Rhea: write my Readme.md (done) and finish debugging
• Abhiram: Start writing packages for PARI/GP and SageMath to compute L functions associated to real quadratic fields.
• James: I think learn how to use CAMB/CLASS to model Pk? Talk to jia about this and results.
• Peter: Speak with Jia, plan out next steps
• Hideki: I try another UNet network to train the relation between galaxy and DM distribution.
• Antoine : make plots, write scripts, check how to make things faster for the analysis - ML for neutrino analysis, replace part of the original maximum likelihood code
• Thomas: read a paper on quantum gravity lab experiments.
• Albert : write the bibliography of my report yay - done!

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

James

Yona

Had a discussion with Jingjing on the cross-correlation: it’s not that straightforward… but exciting perspectives!

​

Made a list of cool things to do in Lyon:)

https://docs.google.com/document/d/1fn2nS1RM9Nsiv_SCwj3NR8H2kKIflBcHQY9-jtiSUrA/edit?usp=sharing

No more cake!

Discuss - future improve of hack format

• Once every few weeks - e.g. sam introduced reinforcement learning
• Some dedicated discussion on technical details
• Data format (dealing with big data, hdf5, pandas)
• Collect a list of notebook (jupyter notebook)
• Jupyter book: https://jupyterbook.org/en/stable/structure/toc.html
• Things to learn
• Foundation model
• Transformer
• Diffusion model
• Energy based
• multi -model
• Parallel computing on idark

CD3 Hack Friday

Jun 16, 2023

Your goal this afternoon

• Jia: Read Daniela’s draft on baryons with HSC Y1 data
• Junjie: Discuss with Jia for optical2radio, write technical notes -> discussed, introduction finished. Bonus: bug fixing of reconstruction codes.
• Cesar: Hyper-Kamiokande deciding near detectors arrangement -> doing sensitivity studies.
• Joaquin: Help Peter with his research project. Finish a draft.
• Yona: ask Jingjing some questions + write 2 more pages of report :) → done, +1.5p written!
• Antoine : keep reading some other guy’s code (pipeline to convert C++ files into ML usable datasets + event selection visualization), try to run some parts and link it to my project
• Valentin : some equations to do (Klein-Gordon and Friedmann equation), and previous equations to write in overleaf
• Jingjing: talk with Toshiki about intrinsic alignment of galaxies (IA) power spectrum and finish revising IA guide draft
• Linda: fix code for initial conditions for clustering dark energy simulations
• James: Talk with jia about plots, implement a new feature in code/debug, finish slides for next week’s astro seminar
• Julien : revise GR because I need to retake my exam:/ (also write function to estimate noise after component separation in fgbuster)
• Peter: try to understand Valeri’s code on GW x galaxy clustering for H0

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Cesar

Predictions agree both for nominal model & for model tuned based on unknown model data (good starting point).

CD3 Hack Friday

Jun 9, 2023

Your goal this afternoon

• Cesar: Install new version of highLAND, run fake data study…
• Katya: make fit with Galight
• Joaquin: Prepare referee report (with chatGPT).
• Albert : fine-tune GNN, write a code to interpret the resulting model
• Zhaoxuan: fit a disk model with 3D Barolo
• James: FIgure out a bug/misunderstanding in the dataset i’m using. If no sufficient progress could be made without help, I will read a couple papers for more background knowledge
• Antoine : debug pointnet code
• Andrés: Make some plots and understand how to run the reconstruction in WAGASCI (detector)

​

CD3 Hack Friday

Jun 2, 2023

Your goal this afternoon

• Jingjing: write the netflow code for PFS cosmology fiber assignment
• Yona: make the slides for the Hongo StudentFest + build cats for Mr<-22
• Andrés: Understand how reconstruction code works for WAGASCI neutrino detector of the T2K experiment.
• James: Finish writing a script for running NEXUS+ on the Quijote simulation set, run it on 512^3 resolution. Also make density fields that have CDM + massive neutrinos.
• Cesar: Play a bit with toy-model binned likelihood fitter (python based).
• Antoine : try to run the WatChMaL (ML framework to reconstruct neutrino interactions in water cherenkov detectors) data production code on new events, finish monday’s presentation
• Albert : read some papers and make some slides yay
• Beatriz: try to install GENIE (neutrino interactions simulator).
• Jules: create more sophisticated probability distribution to train my model
• Akira: fix problem in FastPM simulation

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Figures missing but getting there :)

CD3 Hack Friday

May 26, 2023

Your goal this afternoon

• Jingjing: work together w/ Yona to clean up the code and try to get correct results - did density sample
• Thomas: respond to collaborator, begin processing stellar sample to derive density template. - done!
• Derek: paper proof
• Sam: read ai ethics papers + respond to email. If time continue develop evals code. - talked with abhi
• Jia: talk to Xiangchong, write 1 paragraph in half dome code release paper (fyi - globus may happen, thx to hideki’s help)
• Valentin : go deeper into Brane inflation literature
• Yona: fix fgot, z retrieval in shape cat building, then compute IA
• Joaquin: Work with Daniela in MCMC code for Minkowski Functionals – work in progress
• Daniela: run baryons code, add references paper. Work with Joaquín in MCMC code for MFs
• Antoine : see what people have done with timing in cherenkov detectors. Learn about CNN architectures for data with a time dimension - idea could be useful for different scenario
• Julien : Harmonic component separation in fgbuster : fix some mistakes. Plot residuals and then cosmological likelihood. Verify harmonic component separation with beam, compare with basic component separation.
• Kosuke : Run EMT code to estimate the effective transmittance of pyramid-structured alumina used as a AR filter in SO - success!
• Hideki: Run DisPerSE for the mock 4MOST galaxy catalog to make a filament catalog - found filaments!
• James: finish resolution timing/testing of cosmic web code CosmoMMF, laying groundwork for the flagship analysis of our Massive Neutrinos Cosmic Web Project. Will need to compare computed power spectra for comparison. Look into halo catalogues of Quijote - learned from Hideki Redshift space distortion, maybe a new project idea?
• Albert : fine-tuning of GNN, write some latex, learn about random forest…
• Katya: learn how to use Python’s astroquery for bulk queries to NED and Simbad → import fail: OpenSSL library issues
• Zhaoxuan: update my figures with HST images released by grizli team

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Jia: xiangchong was unavailable

so started learning game theory

https://www.universityofcalifornia.edu/news/how-two-matchmakers-won-nobel-prize

Julien

Hideki (1.7 million galaxies -> 37,000 filaments)

Kosuke

CD3 Hack Friday

May 19, 2023

Your goal this afternoon

• Junjie: read about GEANT4 “physics list” for hadronic interactions - done
• Baptiste: port code to idark, use it to produce a graph showing the effect of miscalibration of the polarization on foreground removal and estimation or r - started coding by stopped by coffee
• Akira: try GELU & SiLU, rewrite SSIM loss function, try chat gpt plugins (super resolution imagine)
• Jia: read Derek’s note on neutrino sims - half way, then chatted about HSC data
• Derek: write summary paragraph of paper - did valuable things..
• Daniela: write paper on HSC Y1 baryon - wrote some
• Leander: write paper (different from above) pdf baryon - globus..
• Sunao: migrate hsc 3x2pt likelihood code to CosmoSIS - wrote some code, also discussed with student on jax
• Katya: run galight with different supersampling factors to check if this option makes big difference. → Almost no change in fit results.
• Boris: Re-write my code for follow-up work clumpy galaxies at high-z - nothing was done
• Thomas: address co-author’s feedback on referee response. - done
• Linda (visiting til Aug, MPA): task for Euclid likelihood computation
• Valentin : compute field at cross-horizon for the Brane Inflation model - didn’t do that, debugged
• Albert : find a way to make the training loop faster, fine tune the parameters
• Yona: finish the IA section of my thesis and fix the cz bug → all done, waiting for the plots :)
• Hideki: Sensitivity estimate of relativistic SZ for LiteBIRD - talked to IT about globus
• Julien : repair my bike…- done
• Xiangchong (visiting til July, CMU): galaxy shear measurement - automatically optimize hyperparams, jax issue, need better parallelization
• Valeri: referee report - updated 1 figure
• Clement: nice plot and paper writing

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Abhiram’s slides

Pitched 3 problems, generated some interest.

Please note that since I will move on from IPMU soon, it is best to get started asap.

CD3 project initiation

Abhiram Kidambi�

I use physical insight to develop new mathematics. ���Often, this developing new mathematics is restrictive due to the infinitude of mathematical possibilities. ����One could use computational techniques to either search for patterns, or a counter-example. Patterns in mathematics usually to be less complicated, often leading to a reasonable sanity check of a conjecture. �����Outline three major problems on which I work, and for which the use of ML techniques could be/are proving to be a game changer.

Problem 1:��Singularities of Feynman Integrals

Path Integrals

Everything that can occur, will occur.

Classical principle of least action��Unique path that minimizes action

Quantum principle of least action��Weighted contribution from all paths that minimize action

Feynman PhD thesis

1. All possible interactions contribute, but the more complicated the path the less it contributes

2. Total contribution (TC) should be finite

3. TC is a power series in physical coupling strength

4. At a given order in the power series (loop order), the set of all possible physical processes can be depicted by a cartoon, called a Feynman Diagram

Question: What is the probability that A -> B, where A and B are sets of particles?

5. Each diagram is basically a pictorial description of an integral known as a Feynman integral. There are “Feynman Rules” to convert diagram <-> Integral.

Paths = virtual particles & interactions

Problem: This makes no sense! Where do such integrals even come from?

Can we predict structures underlying the singular structure of the Feynman integral from some minimal amount of graph data?�

Also related to Mandy’s talk from the CD3 opening symposium.

Feynman graphs <-> Singularities

Problem 2:��Primes, quantum chaos, and beyond

Zeroes and special values of zeta functions and L functions

The Riemann Zeta function is one of the most beautiful functions in mathematics.�Key function in (probably) the most important unsolved problem in math.

When is

Conjecture: If then

Either or

It has been observed that the patterns and statistics of zeroes that lie on the critical line seem to obey some random quantum chaotic system. ��Can we spot patterns that exist between a “System of Zeroes”?

Problem 3:

​

Classifying solutions to string theory�(the Landscape)

“Hide away” 6 dimensions (conjectured internal dimensions of space). How you hide away 6 dimensions tells you what you get out in 4d.

Classifying solutions of string theory = Classifying the 6 dimensional internal spaces (manifolds).

Fundamentally manifolds are just spaces that locally look like as if they are flat.

There are many examples and types of such Calabi-Yau manifolds.

​

But a particular class of them (which are particularly interesting to study) can be constructed on a computer by studying polytopes in higher dimensions and “slicing” them *appropriately* (if possible). (My PhD Thesis)

Problem: The number of ways can you slice the polytopes depends on the number of vertices of the polytope and the number of polytopes. “Search space” becomes too large.��So the novel hope is to identify patterns in how to slice these polygons to construct Calabi-Yau manifolds.

The largest example of these 6d spaces are called Calabi-Yau manifolds.

What mathematicians want from ML/AI?

1. Can AI/ML models learn formal mathematics?��2. Can AI/ML models prove/disprove known results/dis-results?��3. Can AI/ML models verify new mathematical statements?��4. Can AI/ML models generate new mathematical statements?��

Computational mathematics and physics can flourish

1. Classification of Feynman graphs
2. Study of zeroes of special arithmetic functions, like the Riemann zeta function
3. Finding viable solutions to string theory
4. Determining symmetries (automorphisms) of objects (functions/matrices/geometries) from minimal mathematical data
5. Constructing quantum error correcting codes�

What *I* want from CD3?

1. Collaborative workflows where physicists & mathematicians can work together�
2. Learn how to use AI and ML tools more efficiently and effectively�
3. More activities (lecture series, working groups, …) �
4. More grant applications..?

If you are interested in any of these topics (learning or contributing) get in touch with me.

Interesting problems to get started

1. Merten’s conjecture: Can one write a model that can come up with the Merten’s conjecture? (Technically disproven but still..)

​

• Can we find patterns in size of orbits in the hail stone problem?�
• Coefficient of modular forms via shooting method?�
• Figure out automorphisms of small matrices or simple shapes.

End of AK’s slides

CD3 Hack Friday

May 12, 2023

Your goal this afternoon

• Katya: research why galight doesn’t work with HSC pdr2. - succeeded
• Yona: fix the bug in the computation of wgp (not enough points to compute? Inadequate patching?)
• Julien : Read paper and rewrite equations :)
• Junjie: Study and develop WCSim
• Thomas: set up and run codes on idark, then create point source mask. - worked!
• Leander: make plots
• Valentin : implement BI model in CLASS primordial module
• Derek: emails
• Jia: check cold dark matter in CAMELS
• Boris: Hopefully tackle the final issue of my paper - tackled thanks to the ‘refreshments’
• Daniela:FINAL(?) mcmc, slides
• Connor: Finish subsection of paper discussion (for mock HSC image release, mini mergers) - finished subsubsection

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Katya

Connor was right, the problem with PDR2 data is the reference for zero-point.

This reference is something called ‘FLUXMAG0’ in the header of the cutout fits file.

FLUXMAG0 = fits_file_header['FLUXMAG0']

For the same object in PDR2 and PDR3 this number is different.

PDR2 'FLUXMAG0' = 0.0

PDR3 'FLUXMAG0' = 63095734448.0194

Since zero-point is calculated as:

zp = 2.5 * np.log10(FLUXMAG0)

I have -inf for PDR2.

Leander

Jia

Julien

Yona

• Semi-fail: o points → 3 points
• Unit incoherence somewhere ?
• Worked on report instead :)

Daniela

Power spectrum

Thomas

CD3 Hack Friday

April 28, 2023

Your goal this afternoon

• Antoine :-) Try out new losses & regularisations for my model ; explore hyperparameters (probability very low)
• Julien :O remove noise and instrumental systematics from transmittance data of alumina flat disks. See if better accuracy for determining the loss tangent + finishing FGBuster exercice and if time, think about how to change the spectral likelihood
• Valentin :/ establish slow-roll parameters for the Brane Inflation model
• Jules:D Test metrics for a bayesian neural network
• Rhea: Finish my GUI (pretty high probability finishing today)
• Cesar: developing new python-based “package” of drawing tools to visualize ROOT-based analysis output.
• Jingjing: talk with yosuke about our current research projects
• Sam: Implement and test model saving / loading and resuming training
• Yue: Checking the modification to the pipeline of CosmoSIS (for Primordial Non-Gaussianity)
• Boris: Is wavelet decomposition useful for JWST galaxy-morph analysis?
• Thomas: get data reduction scripts running. → responded to referee instead
• Akira: read the review paper about neutrinos
• Jia :p train a NN based on CAMELS simulations
• Daniela: ran many chains
• Leander: galaxy - halo parameters

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Julien (NB : plots are wrong! :( correcting that asap))

Fragmentation in high-z galaxies

​

-Boris

Dense NN on TNG power spectrum - Jia

Goal: predict the power spectrum (~30 numbers) from 6 cosmological parameters. Tried:

• Different parameter
• Updated code from stax to flax (in jax)
• No improvement….(and some artifacts..?)

To try:

• Print gradients (chatGPT can write for me)
• Leakyrelu
• Adding k to input
• Learning rate scheduler?
• Deeper/wider network?
• Dropout?
• There must be something obviously wrong…

https://github.com/liuxx479/cosmo_dep_feedback/blob/master/NN_jax.ipynb

CD3 Hack Friday

April 21, 2023

Your goal this afternoon

• Cesar: play with (auto)GPT
• Antoine : Make some code to plot the results from my model
• Boris: Find a simpler way to visualize my results (too many plots right now)
• Junjie: Explore better python-ROOT interface -> failed
• Derek: Talk to Adrian -> convinced him about the neutrino simulation idea)
• Jingjing: discussion with Minh
• Yona: Finish and clean treatment of sample 18 -> worked!
• Albert : fine tune the learning_rate, etc -> not working suddenly
• Hideki : discussion with Albert -> done
• Jules: find a metric to measure the noise reproduction accuracy of my model
• Claudia: Deep learning 101 with ChatGPT
• Sunao: try Francois’s ML code for my microlensing study.
• Yue: Play with DES data using TreeCorr; tidy up the notes and tools in the past weeks
• DLed data, customizing TreeCorr, to write a job script before running on iDark; tidy up-> failed. Discussed with Giovanni, etc.
• Baptiste: finish proposal
• Leander: think about learning galaxy halo connection
• Jia: discuss with Adrian on kSZ -> Ended up discussing with Derek and Jingjing about neutrino simulations
• Rhea: Modify GUI code to annotate images with points and stop moving the image when moving the points (half success)

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Cesar: (playing with AutoGPT) https://github.com/Significant-Gravitas/Auto-GPT

Auto-GPT concept (use GPT to run GPT…) (define up to 5 goals, let GPT-4 plan tasks to achieve goal and spawn GPT agents to do them), Can: search the internet, read files, save files…

​

​

Running several (~10) tests costed

Leander: learning galaxy halo connection

(Halo Occupation Distribution = HOD)

Idea: maybe the uncertainty in the galaxy halo connection is relatively low dimensional, but with typical HOD/… heuristics we’re missing the simplicity because we don’t include as much physics as is present in a hydro sim

​

Moving parts:

1. Figure out a way to parameterize a mapping from halo properties to a set of galaxies. My idea: autoregressive model (input halo properties & all more massive/luminous member galaxies).
2. Marginalize over uncertainty. My idea: parameterize through a relatively small set of marginals and figure out a way to deform the distribution in a “minimal” way under the constraint that the deformed distribution matches the required marginals. The marginals need to be chosen judiciously based on physical intuition, by going to different hydro sims we can check if they capture the complexity and get ideas for prior ranges.

Antoine : plotting ok

Boris

Jia - discussed with (learning from) Derek new ways of running neutrino simulations

Sunao: simulation based inference for microlensing data

Goal: training P(model params | light curve data)

flux

time

Marginalized posterior

Red: true, Blue: trained posterior

Failed to train, maybe my modeling is bad.

But now I got how to use ML.

You can also take a look at Francois tutorial.

CD3 Hack Friday

April 14, 2023

Your goal this afternoon

• Jingjing: prepare the galaxy/halo shape data in TNG300 for colloborators
• Valentin :) continue exploring the CLASS documentation
• Chenze ; ) Literature review on the accretion shock heating in simulations & figure out how to identify shock in SPH simulation
• Derek: paper revisions
• Jules :D adapting the training loop for a deep generative bayesian network
• Antoine :) write new analysis scripts in Watchmal
• Julien Tang :) finish reading papers on maximum likelihood algorithms for probing CMB
• Shenli: See what is going on here.
• Yona: accurately assign redshift to random catalogues + write some doc :)
• Junjie: learn Bayesian generative model; read CMBxML paper
• Boris: Find energy to start writing my paper
• Albert : correct the training loop of my code
• Thomas: writing ~30% of my paper method, then reviewing another paper.

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

Jingjing: code successfully run on a cluster failed on a new cluster – good documentary tracking package installation needed

Yona: works! (with different method)

CD3 Hack Friday

April 7, 2023

Your goal this afternoon

• Yona: cross match LSS galaxy survey with SDSS shape catalogue, then plot Wcc for one volume-limited sample.
• Chenze: Complete the routine of mock spectra, and attempt some parallelization
• Clement: find where to modify CLASS for implementation of PCA method for reionization
• Shunichi: make plots of cosmological priors for dSph parameter estimation
• Baptiste: debug my power spectrum estimation
• Xuheng: Make a pip release for Galight new version, prepare for CD3 symposium
• Boris: Try to get my head around morphological parameters from JWST galaxies (basically make a ton of plots)
• Ilya: write MCMC module for FRB stuff, estimate DM_igm for FRB20180906B, drink 2 beers
• Yue: install and configure the environment for CosmoSIS on iDark, drink 1 beer.

• Jingjing: debug the calculation of subsample ratio of my cluster IA in the mock
• Sam: Prep for the CD3 symposium
• Antoine : Implement Sinkhorn AE in WatChMaL
• Cesar: Discuss available datasets for ILANCE projects.
• Derek: CD3 symposium literature review
• Albert : parallelize data importation
• Thomas: writing introduction from existing thesis.
• Zhaoxuan: try to find the clusters in my catalogs
• Hideki: Summarize ML paper related strong lens for CD3 symposium
• Valentin and Julien : get familiar with FGBuster, read paper
• Daniela: check emails, read paper, debug code, 1 beer
• Jules: improve the complexity of a ring generation dataset
• Junjie: Discuss plan for paper review @ CD3 symposium; brainstorm ideas to develop DL for neutrino physics

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

• Yona : didn’t work (empty files) … >.<
• Thomas: almost finished the paper introduction (~ 1 paragraph left).
• Baptiste: the scaling of the polarised beam wrt the "temperature" beam after component separation might be the issue.
• Clement: complete failure, try again next time
• Chenze: The script works well, but need some optimization to reduce the memory footprint => benchmark

​

• Yue: Finished. Failed to drink 1 beer→coffee instead.

Boris

CD3 Hack Friday

March 31, 2023

Your goal this afternoon

1. Connor: Get the fraction of merger-triggered Star Formation with IllustrisTNG
2. Sunao: prepare for the CD3 symposium with Xuheng, and prepare for HSC webinar (4/3 midnight!)
3. Daniela: Create HSC weak lensing mocks based on BAHAMAS sims
4. Derek: work on uniform velocity integral equation/initial conditions (for neutrinos)
5. Xuheng: JWST filter calibration and error budget
6. Jia: finish leander’s draft section “Method” (maybe also “results”)
7. Leander: either batched emcee or finally start with pdf baryons paper (nested sampling<-sunao)
8. Sam: Understand/reproduce Fig 5 of https://arxiv.org/pdf/2303.08774.pdf Japanese GPT performance
9. Akira: check if redder member stars of Sextans have different proper motion compare to bluer ones. (HSC SSP)
10. Boris: Find correlation between bulge formation and clumps in the young Universe that is being revealed by new JWST data (and reach 8% blood alcohol level for scientific purposes).
11. Angela: Learning the 1pt and 2pt measurements for DES catalogs, understanding how the masks and systematics maps work (treecorr)
12. Rhea: Using PySimpleGUI to build GUI used for Super-K photogrammetry feature detection/labeling
13. Junjie: Reading and understanding the new project about optical to radio galaxy mapping; write internal instructions for T2K data reduction

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

End of hack report

• Connor: Get the fraction of merger-triggered Star Formation with IllustrisTNG -> see figure
• Sunao: prepare for the CD3 symposium with Xuheng, and prepare for HSC webinar (4/3 midnight!) -> created repo
• Daniela: Create HSC weak lensing mocks based on BAHAMAS sims -> see fig
• Derek: work on uniform velocity integral equation/initial conditions (for neutrinos)
• Xuheng: JWST filter calibration and error budget -> found paper, 0.06 magnitude, reasonable (budget is 0.2 magnitude)
• Jia: only finished 1 subsection of the section.. 2 more subsections 2 go..
• Leander: either batched emcee (nested sampling<-sunao) -> boring, gave up�or finally start with pdf baryons paper-> turned out to be less boring
• Sam: Understand/reproduce Fig 5 of https://arxiv.org/pdf/2303.08774.pdf Japanese GPT performance
• Boris: Began running the estimation of bulge to disk ratio of galaxies with MCMC generated error-bars (also achieved <6% blood alcohol content)

End of hack report

• Akira: check if redder member stars of Sextans have different proper motion compare to bluer ones. (HSC SSP) -> project aborted, instead wrote referee report with chatgpt??
• Angela: Learned how to open Jupyter notebook on idark from Connor Bottrell, read some codes of DES lss measurements, made my first ngal plot
• Rhea: Using PySimpleGUI to build GUI used for Super-K photogrammetry feature detection/labeling
• Junjie: Reading and understanding the new project about optical to radio galaxy mapping; write internal instructions for T2K data reduction
• Daniela: wrote code for hsc mocks, testing non gaussian statistics

Connor

daniela

Junjie

• Played with the data from 2020ApJS..247...53K (Radio sources associated with Optical Galaxies and having Unresolved or Extended morphologies (ROGUE).
• ​

Maybe 2d histogram instead of scatter plot

​

Check multiple bands B-V, R-I, G-I to get SF

​

Match to SDSS DR7 galaxies

Sam

Tried to understand this plot:

Cost to reproduce >~ $3000

Angela’s note

One way to use jupyter notebook on idark (from Connor Bottrell):

1. Install jupyter notebook, for example, using conda in your idark environment
2. In your local bash profile file, add this line: alias sshuttle_idark="ssh -NfL 8889:localhost:8889 user@idark.ipmu.jp"
3. “jupyter lab --no-browser” on idark head node
4. Open a new local bash shell, and type

“sshuttle_idark”

• Copy and paste the jupyter lab link from idark to your local browser

Boris

Each beer today had 14g alcohol which contributes to 3% BAC for a 68kg male

https://www.sciencedirect.com/science/article/pii/S1053810012000037

https://xkcd.com/323/

CD3 Hack Friday

March 24, 2023

Introducing Hack Fridays (Jingjing, Cesar)

• The goal: tackle data and coding challenges together & accelerate your project(s)
• The format: Friday 1-5pm seminar room C (3F)
• 1pm-1:20pm each participant describes their goal of the day (30s-1min/person);
• 1:20-4:30pm find a comfy place (the seminar room, the balcony, the 3F open space, anywhere!) to code with your buddy(ies) or alone; discuss your bugs/new ideas freely with anyone; we will have experienced coders to support basic programming issues;
• 4:30-5pm reconvene and report back your accomplishments (or failures..); get feedback from others
• What to bring:
• a laptop
• a small task you plan to accomplish over an afternoon. Time is short, so be realistic (e.g. fix a bug, write a few functions, or write the abstract of your paper; do NOT expect to finish your PhD thesis there!)
• An open mind!
• Who can/should join?
• Anyone who codes. We want this to be a minimal burden to your already busy work/life, so feel free to join and/or leave anytime.

Your goal this afternoon

1. César: (C++, python) Read Wagasci reconstruction code.
2. Daniela: (python) Compare methods to average maps.
3. Sam: (python) Learn how to use langchain
4. Yona: (python, fortran) successfully run Treecorr and get some output (not necessarily with the right data)
5. Xuheng: (python) Testing the SED fitting accuracy for two band photometry
6. Abhiram: (python + Mathematica) Improve code to find patterns in mod p q-series. Get to p = 11. (Something should go wrong there, find out why)
7. Derek: (fortran, python) solve some integral equations
8. Albert : (python) learn the basis of pytorch, try to understand a code written in it
9. Claudia: Analyze where signal is coming from (python)
10. Hideki: Run test code with Albert
11. Leander (C, C++, python): try <z^2 kappa> ??? Otherwise baryons in pdf paper
12. Sunao: (python) Run several MCMC chains for HSC weak lensing project (& paper writing).
13. Jingjing (python, C, C++): revise the IA cluster paper; test the cluster IA
14. Akira (python): read fugaku documentation for simulation
15. Jia: read leander’s HSC Y1 lensing pdf paper draft (uses python usually)

End of hack report 4:30pm

Please share 1 slide of your accomplishment / failure..

End of hack report

• César: (C++, python) Read Wagasci reconstruction code -> Moved forward with reconstruction code documentation.
• Daniela: (python) Compare methods to average maps.
• Sam: (python) Learn how to use langchain
• Yona: (python, fortran) successfully run Treecorr and get some output (not necessarily with the right data)
• Xuheng: (python) Testing the SED fitting accuracy for two band photometry -> simulation performed with result summarized here
• Abhiram: (python + Mathematica) Improve code to find patterns in mod p q-series. Get to p = 11. (Something should go wrong there, find out why) ->Unary theta function is a subseries at mod 2!!
• Derek: (fortran, python) solve some integral equations -> solved to better than 5%
• Albert : (python) learn the basis of pytorch, try to understand a code written in it -> coded some thing
• Claudia: Analyze where signal is coming from (python) -> found some pattern
• Hideki: Run test code with Albert -> done
• Leander (C, C++, python): try <z^2 kappa> ??? Otherwise baryons in pdf paper -> doesn’t work (some extra contamination)
• Sunao: (python) Run several MCMC chains for HSC weak lensing project (& paper writing). -> NA
• Jingjing (python, C, C++): revise the IA cluster paper; test the cluster IA
• Akira (python): read fugaku documentation for simulation -> read doc, read spin 2 cnn paper, now need to learn jax
• Jia: read leander’s HSC Y1 lensing pdf paper draft (uses python usually) -> edited 2 paragraphs…

Daniela

I checked the discrepancies between the two average methods and the impact on the project’s result.

​

Thanks Jia, Leander and Albert for the discussion!

Sam

https://www.ipmu.jp/en/20230324-CD3

Possible improvements

• Music: each person takes turn to DJ
• Pizzas (pizzahut, dominos)
• Redbull
• Start description from 1:30, but can start working/listen to music from 1
• Create a mailing list (after 2-4 weeks)