Project Expo 2025

13:30–15:00, Friday, May 30th, 2025

Lecture Hall, Kavli IPMU

CD3 Year 2 Summary

Jia Liu

Workshops, Conferences, and Schools

​

1. The Quantum Frontier of Machine Learning (March 11, 2025)
2. LiteBIRD Collaboration Meeting (January 7-9, 2025)
3. PFS Collaboration Meeting (January 20-24, 2025)
4. Future of AI Research for Science in Japan (FAIRS-Japan) (December 3-5, 2024, Nagoya University)
5. Probing the Genesis of Super Massive Black Holes (November 18-21, 2024)
6. Focus Week on Primordial Black Holes 2024(November 13-15, 2024)
7. AstroAI Asian (A3) Network Summer School (September 2-6, 2024, Osaka University)
8. LSS Quest 2024 (June 24-25, 2024, Osaka University)
9. Workshop on Galaxy and Black-hole Evolution (June 3-7, 2024)
10. CD3 Project Expo 2024 (May 10, 2024)
11. Baryons in the Universe 2024 (April 8-12, 2024)

PFS Collaboration Meeting

Code of Conduct

CD3 is committed to creating a safe, inclusive, and respectful environment for all members of our community. CD3 members are required to:

Treat each other with respect and professionalism. Discrimination, harassment, and retaliation of any kind will not be tolerated.

Follow the highest standards of academic integrity and honesty. This includes, but is not limited to, avoiding plagiarism, fabrication, and falsification of data. All members are required to comply with Kavli IPMU’s Research Ethics.

Researchers using AI tools should document this use in the methods, acknowledgements, or other appropriate sections in the paper. Do not use AI tools as a credited author on a research paper, as it can not carry accountability.

Housekeeping & Announcements

• “CD3 x astro/math/pheno/CMB/neutrino” seminars (Contact: Leander)
• Your speaker may be funded by CD3 if they work on data science/AI/ML
• Hack Fridays, 13:30-17:00 (Contacts: Katya, Joaquin, Enrico)
• Playground for data-related study groups, project collaborations, and technical discussions
• Ongoing: Numerical Simulation Focus Group, Human-AI Collaboration Project
• Joining CD3:
• Sign up on https://cd3.ipmu.jp/
• ChatGPT Team account
• Funding for workshops, visitors, and travel
• Transition to non-academic career
• Required commitments: engage in CD3 activities, publish with CD3 affiliation, follow the Code of Conduct

Incoming Postdocs

Otavio Alves

Michele Veronesi

Vera Maiboroda

Takafumi Tsukui

Andrew Santos

Luca Marchetti

Rahul Ramesh

Nisha Grewal

3rd Year Focuses

• Science:
• Effective integration of AI into our workflow (ChatGPT Team)
• QFT x ML program
• “Human-AI Collaboration” program�
• Education:
• AstroAI Asian Network: 70 students participated our first summer school in Osaka in 2024. Second summer school will be in Seoul, Korea

AI for

literature review

Joaquin Armijo, Linda Blot

AI tools for literature review

Why use AI Tools?

• Save time and reduce information overload.
• Improve precision in identifying relevant research.
• Enhance clarity and structure of literature reviews.
• Discover interdisciplinary insights more easily.

How do AI tools help?

• Automate searches and extract key papers using natural language.
• Summarize findings across multiple studies quickly.
• Identify trends, gaps, and key contributors in a field.
• Organize citations and create conceptual maps.

What tools are available?

AI-powered software for searching, analyzing, and summarizing academic literature.

Examples (not limited to): ChatGPT, Scispace (formerly known as typeset.io), Research Rabbit and Anara

​

Using ChatGPT for literature and insight extraction

Example Task: “Summarize key ML approaches used in weak lensing.”->

Useful for early-stage scoping, defining keywords, identifying popular methods

Limitations:

• Not a database. Can't cite unless you feed specific papers or provide citations.
• May hallucinate references if asked for citations without grounding.

​

Using ChatGPT for literature and insight extraction

Good level of knowledge. Provides references when needed!

Using ChatGPT for literature and insight extraction

Not only papers, but useful links to different tools (in github). Ml-in-cosmology, CNN_mass_maps

SciSpace for Deep Reading and Concept Linking

Example Task: “Understand methodology sections of several weak lensing + ML papers.”->

Not as conversational as ChatGPT—designed for deep dives, not exploration.

Provides synthesis of information.

SciSpace for Deep Reading and Concept Linking

​

ResearchRabbit for Visual Literature Exploration

Example Task: “Map the key papers and authors in ML + weak lensing.”->

• Great for building reading lists.�
• Shows evolution of subtopics, who cites whom.�
• Alerts you when new articles appear in your “rabbit holes.”

For example: Input the papers from queried by ChatGPT, SciSpace and generate graph (network/timeline)

Combining the Tools in a Literature Pipeline

Suggested by ChatGPT

Summary & Conclusions

• AI tools can significantly accelerate literature review
• Combining these tools creates a powerful and repeatable system for reviewing literature efficiently.
• Helps to manage the overwhelming volume of papers.
• Limitations exist, but can be mitigated by using the tools together and maintaining critical oversight.
• The barrier to entry is low: free or affordable access, intuitive interfaces, and rapid onboarding.�

​

AI for Coding

Jess Cowell, Ben Horowitz

AI

PHYSICS

cd3 expo 2025

AI FOR CODING

HIGHSCORE 2500

WHAT IS AI-ASSISTED CODING?

• Compose code using natural language, using LLMs.

​

• Receive assistance with debugging and intelligent suggestions, powered by AI models created from millions of code repositories.

​

• The emerging trend of “vibe coding” allows users to generate code from scratch using LLMs, even without prior coding knowledge.

WHY WOULD YOU USE?

• **GitHub Copilot Study (2023)**:
• 55% faster at coding tasks
• Increased satisfaction and focus among participants

​

• **Microsoft Research with Codex (OpenAI)**:
• 71% of developers reported productivity boosts
• 88% more likely to explore unfamiliar code

​

• **Stack Overflow Developer Survey (2023)**:
• Over 70% of developers use or plan to use AI tools .

TOPICS COVERED

level 1

level 2

level 3

final boss

AGENDA

prompting

tips and tricks

chatgpt

ide editors

PROMPTS

prompts are so important people even

do 18 hour courses on them!

​

• Be specific: What do you want the code to do?
• Include context: e.g., “make a 5x5 subplot using matplotlib”
• Add constraints: input type, or language
• Paste code or errors: Give the AI something to work with
• Break tasks into steps: “First write function, then plot result”
• Describe your goal: “I want to visualize S₈ vs Ωₘ from MCMC chains”

​

​

HOW TO PROMPT

LEVEL 1: CHAT GPT

• Easy to use, and included with CD3!
• All rounder, can search the internet for examples, of code, voice chat, and has a memory of your previous chats

Can you explain how to use except statements in python?

EXAMPLE 1: NEW CONCEPTS

Here’s a concise 5-line summary of what the code does:

• Initializes cosmologies with different dark energy parameters w0w_0 and waw_a using the CLASS code.
• Defines several functions to compute the linear growth factor D(a)D(a) and growth rate f(a)f(a) using analytic approximations and integrals.
• Computes background cosmological quantities like the Hubble rate and matter/dark energy densities from CLASS for each model.
• Generates plots comparing the CLASS-computed growth factor and analytic approximations across scale factor aa, including relative errors.
• Saves visual comparisons between ΛCDM and varying dark energy models to PDF files.

Can you summarise this code in 5 lines?

<PASTE SOME CLASS CODE SNIPPET HERE>

EXAMPLE: CLASS CODE SNIPPET

Can you explain how it calculates the Hubble rate?

EXAMPLE: CLASS CODE SNIPPET

LEVEL 2: GITHUB COPILOT PLUGIN

An AI assistant plugin for code editors (e.g. VS code, cursor)

Suggests code as you type- like ctrl tab but generates entire lines of code

Trained on open-source code from GitHub

NEW CODING AGENT...? (MAY 19TH)

BONUS LEVEL: TERMINAL INTEGRATION

TIPS AND TRICKS

A WORD OF WARNING

Beware of physics or maths mistakes in your code! They can be subtle...

​

Real life examples;

• Missing a factor of 2pi in a fourier transform
• Missing an exponential factor in an equation
• Not applying suvat equations properly to a game simulation so balls were not bouncing to the right height

​

<Ai makes mistakes>

“When AI Gets the Science Wrong”

DON’T LET IDE EDITORS GET CARRIED AWAY

Real Example:

​

• You request it to neaten your code before uploading to GitHub or sharing with someone.
• All your files are adjusted to meet the highest professional coding standards.
• Your previously functioning code is now broken, and you can’t understand it anymore... But, it looks professional!

IDEs like cursor can rewrite all files in your directory if you let them

DON’T END UP LIKE THIS...

Using AI responsibly

Katya Vovk, Joaquin Armijo

How you use AI

(in writing, coding, translating, reviewing, etc.)

How AI is designed

(by developers, to make systems safe and fair)

Responsible AI

Using AI Responsibly

Privacy & Data Governance

Transparency & Reproducibility

Fairness

Plagiarism by AI. Even with vague description, it generates copyrights protected images

Privacy & Data Governance

Transparency & Reproducibility

Fairness

Privacy & Data Governance

Transparency & Reproducibility

Fairness

Lapushkin et al. (2019)

Request to generate just soda results in certain brand exclusively.

Privacy & Data Governance

Transparency & Reproducibility

Fairness

The Golden Rule

Read the guidelines: every journal, funder, and institution has evolving policies.

If you’re unsure whether an AI use is appropriate, ask:

”Would I be okay justifying this in a peer review or to a grant panel?”

CD3 Project Expo 2025

Stay after for a group photo!

Neutrino programs: SK, HK, T2K

Presenter: Elisa Ferreira

Experts: Patrick de Perio, Mark Vagins, Tsui Ka Ming

Elisa Ferreira

Experts: Mark Vagins, Patrick de Perio, Tsui Ka Ming

Neutrino programs:

T2K, SuperK, HyperK,

Super-Kamiokande=Super-Kamioka Neutrino Detection Experiment

• Located at 1,000m underground in the mountains of Kamioka Town, Gifu prefecture
• Stainless-steel tank, 39.3m diameter and 41.4m tall, filled with 50,000 tons of water, 11,146 photo-multipliers

Science goals:

• Study the properties of solar, atmospheric and man-made neutrinos
• Observations of the universe by neutrino e.g. from supernova
• Search for proton decay

Neutrino scatter off electrons or nucleus, relativistic charged particle

=> Cherenkov light is emitted in a cone shape to the direction of a charged particle, detected by the PMT on the tank’s wall

Gadolinium upgrade (SuperK-Gd)

Enhances ability to detect neutrino by increasing the efficiency of neutron capture on gadolinium, which improves the detector's sensitivity to neutrino interactions

​

• Goal: 0.2% concentration of a �gadolinium compound

Science goal:

• Observation of supernova relic neutrino from the diffuse Sn background

Science goal:

• Observation of supernova relic neutrino from the diffuse Sn background

Gadolinium upgrade (SuperK-Gd)

• Computational needs: 14 Tb/yr raw data
• Monte Carlo (MC) simulations tuned with calibration data to generate observables with which we use to analyze real data
• Trigger rate (number of event candidates detected per unit time) increases significantly with a lower energy threshold. Current: 20 kHz at a threshold of 3.5 MeV. If threshold is lowered, requires efficient real-time parallel online computing capabilities to manage and analyze the data stream.

Enhances ability to detect neutrino by increasing the efficiency of neutron capture on gadolinium, which improves the detector's sensitivity to neutrino interactions

​

• Goal: 0.2% concentration of a �gadolinium compound

Science goals:

• Precision neutrino oscillation measurement
• Search for CP violation

Tokai-to-Kamioka

Neutrino beam created at J-PARC (accelerator)

Near detector: measure neutrino before it oscillates

SK: Measure neutrino beam after it oscillates

Tokai-to-Kamioka

Neutrino beam created at J-PARC (accelerator)

Near detector: measure neutrino before it oscillates

SK: Measure neutrino beam after it oscillates

Science goals:

• Precision neutrino oscillation measurement
• Search for CP violation

Detector:

• Near detector very different than SK
• ND280 detector very close to the detector: more data but needs great precision to control systematics

Science goals:

• Precision neutrino oscillation measurement
• Search for CP violation

Tokai-to-Kamioka

Neutrino beam created at J-PARC (accelerator)

Near detector: measure neutrino before it oscillates

SK: Measure neutrino beam after it oscillates

Detector:

• Near detector very different than SK
• ND280 detector very close to the detector: more data but needs great precision to control systematics

• Event reconstruction for each detector very challenging and expensive
• Statistical analysis for T2K’s for a huge datasets

September 20, 2024

World's largest artificial underground cavity

(45 out of 71 m deep)

​

Public tour June 29! �Space limited, apply here by Sunday night! https://www-sk.icrr.u-tokyo.ac.jp/news/detail/1705

• 8.4 x Super-K
• New high-sensitivity photosensors
• Upgraded J-PARC accelerator facility for neutrino beam

Main science goal:

• 5σ detection (10 years) of CP violation in neutrino sector

Main science goal:

• 5σ detection (10 years) of CP violation in neutrino sector

• 8.4 x Super-K
• New high-sensitivity photosensors
• Upgraded J-PARC accelerator facility for neutrino beam

Need to reduce to <3% for future!

T2K 2021 sys

• 8.4 x Super-K
• New high-sensitivity photosensors
• Upgraded J-PARC accelerator facility for neutrino beam

• Computational needs:
• 35PB of data (including MC simulation data)
• ~8700 cores
• ML methods to improve the quality, speed and resource need

Main science goal:

• 5σ detection (10 years) of CP violation in neutrino sector

Differentiable Detector Simulator (DDSim)

• Traditionally use Monte Carlo (MC) simulations tuned with calibration data to generate observables with which we use to analyze real physics data

​

• Traditionally use Monte Carlo (MC) simulations tuned with calibration data to generate observables with which we use to analyze real physics data
• Idea: develop and use Analytical Differentiable Simulation or differentiable surrogate models for simulation, calibration and reconstruction

Differentiable Detector Simulator (DDSim)

Automation of physics model tuning (via backpropagation)

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

​

Subaru programs: PFS, HSC

Presenter: Tom Melia

Experts: KG Lee, Jingjing Shi, �John Silverman, Masahiro Takada

CMB programs: LiteBIRD, Simons Observatory, CMB-S4

Presenter: Linda Blot

Experts: Tomotake Matsumura, Guillaume Patanchon, Toshiya Namikawa

CMB

Linda Blot

Cosmic Microwave Background (CMB)

Main science goals:

Large scales: primordial gravitational waves from inflation

Small scales:

• CMB lensing: cosmology
• Sunyaev-Zeldovich effect: physics of cluster of galaxies
• Effective number of relativistic species: light relic, such as axion, sterile neutrinos, dark radiation, ...
• Sum of neutrino masses
• Optical depth: reionization

Cosmic Microwave Background (CMB)

Space:

• full sky -> large scales
• No atmosphere -> larger frequency range

Ground:

• Larger telescope -> small scales
• More detectors -> better sensitivity

• Measure polarization of CMB (B-modes)
• Larger number of detectors wrt previous generation
• Challenge: separating foregrounds

Cosmic Microwave Background (CMB)

Space:

• full sky -> large scales
• No atmosphere -> larger frequency range

Ground:

• Larger telescope -> small scales
• More detectors -> better sensitivity

Data Analysis Challenges:

• Huge stream of time ordered data to reduce analyse
• Foreground cleaning
• Delensing
• Many systematic effects to model: atmosphere, ground signal, polarization angles, detector gain…

The Lite (Light) satellite for the study of B-mode polarization and Inflation from cosmic background Radiation Detection

• Uniformly observe the entire sky for 3 years from ~FY2032
• Satellite with single mm-wave telescope
• ~4.5k transition edge sensors across 15 frequency bands between 34 – 448 GHz to distinguish foregrounds
• Cooled to 5 K low temperatures to reduce thermal noise
• Polarization modulation units (PMU) using continuously rotating half-wave plates (HWP) modulate the polarized light and overcover the instrumental excess noise at low frequency where the signal also appears

The LiteBIRD is currently under reformation so the detailed specifications are subject to change

Prototype

@ IPMU 1F

Atacama desert in Chile ~5200m, high and dry: better atmospheric conditions

​

Construction is done! LAT first light in February, currently being calibrated, 3SATs are online and taking data. 3 more SATs are under construction thanks to SO Japan, and SO UK

15m

14m

2 types of telescopes: (SAT 10% of the sky, LAT 50 % of the sky)

• 3 SATs 0.5m lenses, rotating HWP, >30k detectors, FOV:35deg, 0.5° resolution
• LAT: 6m mirror, >30k detectors FOV:7deg, ~arcmin resolution
• 6 frequency bands 30–280 GHz (1–10mm)

SAT

LAT

Large-area survey: light relics

Small-area survey: primordial gravitational waves

1 LAT: 5m telescope, 130k detectors

9 SATs: 75k detectors in total

2 LATs: 6m telescope, 130k detectors each

All detectors cooled to 0.1 Kelvin

XENON

Presenter: Jingjing Shi

Experts: Masaki Yamashita, Kai Martens

Jingjing Shi

Expert: Masaki Yamashita

Science Goal:

• Direct detection of Dark Matter (DM) via elastic scattering off nuclei
• Measuring mass and cross section of DM particle

Rare event:

• Observing from the underground laboratory in Italy (LNGS), 1400 m depth
• Data taking was started from 2021
• More than 5 years operation is planned

​

Looking for dark matter from underground

Detector:

Gas-liquid dual phase Xe Time projection chamber (TPC)

• 8.5 tonne Xenon
• ~500 Photosensor(PMT) (Top and Bottom array)
• Immersed in the gadolinium-loaded water (SK-GD tech)

Gd-loaded Water Tank

LXe Detector

Photosensor(PMT) array

XENON with CD3

⁸B Neutrinos Data Analysis :

Boosted decision tree (BDT) classifiers are developed to distinguish between 8B CEνNS signals and the background events. 

Not only Dark Matter ...

Highlight from XENON

First indication of Solar ⁸B Neutrinos

via Coherent Elastic Neutrino-Nucleus

PRL 133, 191002 (2024)

8B signal

Background

8B signal

Background

Light(S1) and Charge(S2) signal from Photosensors

From left: Kai Martens, Masaki Yamashita

Tianyu Zhu (Postdoc)

Caio Ishikawa (PhD student)

Xiaoxin Wang (student)

Belle II

Presenter: Masaki Yamashita

Experts: Takeo Higuchi

Masaki Yamashita

Expert: Takeo Higuchi

Overview:

• Belle II is located at the SuperKEKB collider in Japan
• SuperKEKB: accelerator in Tsukuba (3km circ)
• Electron/Positron accelerator (7 and 4 GeV)
• Collected data in 2019-2022 and resumed operation in February 2024
• Highlight: The peak luminosity— relevant with electron and positron collisions— has reached 5.1 × 10³⁴ cm⁻² s⁻¹, setting a new world record.

Science goals:

• Probe beyond standard model physics
• CP violation in the physics of B mesons: study

Kavli IPMU members

Takeo Higuchi ... faculty Fanli Zeng ... D3

Fumiaki Otani ... D3

Tomoyuki Shimasaki ... D3

Deven Misra ... M1

19 detector modules produced in IPMU’s clean room, delivered in 2019

Source: arXiv:2402.17260

​

GFlaT: new machine learning tool using graph neural network.

Drastic improvement: 18% better tagging efficiency compared to previous algorithm (see arXiv:2402.17260)

​

Belle II will restart in 2025 Oct!

Silicon Vertex Detector

Data analysis challenge: flavor tagging

i.e. identify from from their disintegration products.

Machine learning theory

Presenter: Guillaume Patanchon

Experts: Simeon Hellerman, Elisa Ferreira, Leander Thiele, Masahito Yamazaki

Deep learning principles

Physics informed ML: QFT/ML

Euclid, LSST

Presenter: Masahito Yamazaki

Experts: Linda Blot, Masahiro Takada

COSI

Presenter: KG Lee

Experts: Tad Takahashi

COSI (Presenter: KG Lee; Expert: Tadayuki Takahashi)

• COSI (Compton Spectrometer and Imager) is a next-generation MeV gamma-ray space telescope, scheduled for launch by NASA in 2027 as a SMEX mission.
• Addresses the “sensitivity gap” in the sub-MeV to few MeV range—critical for both astrophysics and particle physics.
• Uses germanium semiconductor detectors with Compton scattering principles for precise gamma-ray imaging and energy resolution.

Technical Highlights

• Advanced detection: 16 liquid-nitrogen-cooled, double-sided Ge strip detectors.
• High precision:
• Energy resolution: < 6 keV at 511 keV
• Angular resolution: <4.1° @ 511 keV; <2.1° @ 1.157 MeV
• Full-sky coverage daily in survey mode.
• Background rejection: Active shielding with BGO scintillators and BTO (Background and Transient Observer) using NaI(Tl).

​

Scientific Goals – Annihilation Lines & Nucleosynthesis

• 511 keV line: COSI will map this key signal of electron-positron annihilation, helping identify positron origins (e.g., radioactive decay, dark matter).
• Nuclear gamma-ray lines:
• ⁴⁴Ti (1.157 MeV) → young supernova remnants
• ²⁶Al (1.809 MeV), ⁶⁰Fe (1.173/1.333 MeV) → stellar nucleosynthesis and galactic chemical evolution.
• Exceptional energy resolution allows for fine spectral studies.

Multi-Messenger Astronomy

• COSI covers an unexplored MeV window, complementing radio, X-ray, and TeV observations.
• Detects polarization of gamma rays from:
• Gamma-ray bursts (GRBs)
• Crab Nebula
• Cyg X-1
• Key to understanding radiation mechanisms, geometry, and cosmic-ray origins.
• Capture prompt gamma-ray signals from next SN1987A-like nearby supernova.

Exploring DM

• Probes wide mass ranges of DM candidates:
• ALPs (axion-like particles) via polarization signals
• WIMPs (TeV-scale) via MeV-GeV gamma-ray spectrum correlations
• Primordial black holes via evaporation gamma rays
• MeV-scale sterile neutrinos and feeton DM via decay signatures (→ 511 keV lines)
• Enables multi-channel testing of decay, annihilation, and interaction models.
• IPMU Involvement: Tom Melia, Shigeki Matsumoto

ELSI

Presenter: Leander Thiele

Experts: Hiromi Yokoyama

Social Science (ELSI)

Expert:

Hiromi Yokoyama

Presenter:

Leander Thiele

ELSI = Ethical, Legal and Social Implications

more negative attitude

General method:

– conduct online surveys (careful about biases!)

– identify predictors, models

– measure ethical stance of public

Measuring ethics

1. Preconventional morality
1. Obedience and punishment orientation
2. Individualism and exchange
2. Conventional morality
• Good interpersonal relationships
• Maintaining the social order
3. Postconventional morality
• Social contract and individual rights
• Universal ethical principles

1. Risk & Trust in Science
2. AI & Climate
3. Women in STEM
4. Science, Technology and Society Barometer

AI ethics

AI ethics