Astro AI with Fugaku

@ University of Tsukuba, Tokyo Campus

11-12 Sep. 2023

https://fugaku100kei.jp/fugaku/promoting_research.html

​

Keyword is AI!

Day 1 (11 Sep.)

​

10:30 - 12:00 Keynote talks (30 min incl. Q&A)

Yusuke Iida

Jingjing Shi

​

12:00 - 13:30 Lunch break

​

13:30 - 13:50 Self introduction (~1 min per person)

​

13:50 - (17:00) Hack

​

Day 2 (12 Sep.)

​

10:00 - 12:00 Research talks (15 min incl. Q&A)

Hiroyuki Masaki

Yuta Asahina

Akira Harada

Keiya Hirashima

Kana Moriwaki

Akira Tokiwa

Armijo Joaquin

Hideki Tanimura

​

12:00 - 16:00 Free lunch + Hack

​

Feel free to ask questions anytime!

• Network: eduroam.
• Slack: join #workshop_fugakuai_sep23 for communication in hack.
• Please prepare a self-introduction slide.
• Eating and drinking are not allowed in this room.

​

Self Introduction

1 min per person

​

Please put your name, affiliation, and photo.

Kana Moriwaki

• Assistant Professor (the University of Tokyo, RESCEU/UTAP)
• Research interest:
• High-redshift galaxy formation, line-emitting galaxies
• The large-scale structure of the Universe
• Cosmic reionization
• Recently I developed a conditional GAN for signal reconstruction from large-scale line intensity maps.
• I am now interested in new ML models such as diffusion models and Transformer as well as explainable AI and physics-informed AI.

http://www-utap.phys.s.u-tokyo.ac.jp/~moriwaki/

Yusuke Iida

• Associate Professor (Niigata University)
• Research interest:
• Solar dynamo (mainly observation)
• Space weather (prediction models for solar flare and coronal hole)
• I am recently working with non-solar physics field researchers in deep learning modeling using image data (galaxy, meteorology, glacial science, biophysics, agriculture, medical science…).
• My recent interests are symbolic regression technique using AI based method and diffusion model.

Akira Harada

• Special Postdoctoral Researcher (RIKEN iTHEMS)
• Research interest: Supernova
• theoretical model of explosion mechanism
• analysis of multimessenger signals from supernovae
• My dream is to construct AI assisted supernova simulation that achieves high fidelity (accurate neutrino transport/hydrodynamic turbulence) with low numerical cost. I’m also considering to apply AI to supernova neutrino analysis.
• My recent interests are diffusion model (for simulation) and AI-supported symbolic regression (for neutrino analysis)

Hideyuki Hotta

• Professor (ISEE/Nagoya University)
• I have just moved from Chiba University this April.
• Research interest:
• Solar/Stellar interior and surface phenomena
• High Performance Computing
• I mainly depend on my student (Masaki-kun) regarding AI project.

Keiya Hirashima

• PhD student (the University of Tokyo)
• Research interest:
• Supernova feedback and galaxy formation
• Surrogate modeling
• Have applied CNNs to surrogate modeling for supernova simulations toward high-res. sims. of the MW galaxy.
• Some other projects…
• GNNs/transformers for particle-based surrogate modeling
• Foundation model to describe the morphology of galaxies

Hanchun Jiang

• Master student (the University of Tokyo, UTAP)
• Research interest:
• CMB polarization
• Dark matter
• Used to apply the CNN model for searching for M dwarf flares from the observed light curves
• Mainly interested in explainable AI

Yuki Kambara

• Master student (the University of Tokyo)
• Research interests
• Planet formation / Planetesimal accretion from a ring
• N-body simulation
• Very new to AI / ML application
• No experience of application
• Interested in
• accelerating simulation
• analysis / modelling of simulation results

Akira Tokiwa

• PhD student (Kavli IPMU)
• Research Interest
• Galactic Archeology
• cosmological simulation
• On Fugaku:
• run N-body simulation
• Spherical Super-Resolution & Diffusion

Ken OHSUGA

• Professor (Univ. of Tsukuba)
• Research interest:
• Black hole accretion flows and outflows
• Radiation magnetohydrodynamics simulations / Radiation transfer calculations
• Recently, I have been working with Asahina-san on machine learning for radiation transport calculations.

Simulation results:

[left] Black hole accretion disks and jets

[right] Photon ring

Akihiro Inoue

• PhD student (Univ. of Tsukuba)
• Research interest:
• Accretion flows onto compact objects, especially magnetized neutron stars
• General Relativistic Radiation MHD simulations
• AI/ML
• No experience of application
• Interested in accelerating simulations

Simulation results (right panel) :

accretion flows around the magnetized neutron star

Jingjing Shi

• Postdoc at Kavli IPMU
• Research Interest

- Large-scale structure cosmology: galaxy clustering, intrinsic alignment, weak lensing

- Galaxy formation and evolution: analysis using cosmological hydro-dynamical simulations

• ML experience: galaxy-halo connection

​

Joaquin Armijo

Denoiser weak lensing (convergence) map

• Joaquin Armijo, Kavli IPMU (CD3), Postdoc.
• Cosmology, large-scale structures, halo-galaxy connection, modified gravity.
• Machine learning for cosmology.

​

Tian Qiu

• PhD student (Kavli IPMU)
• Research Interests:
• Galactic Archeology
• Proper motion of stars
• White Dwarfs
• AI/ML: I am interested in coding, but I have no experience to apply the AI to research.
• To find structures in the Milky Way
• Astrometric measurements

​

• Cosmological simulations, AI/ML
• Cross-correlation between CMB and Large-scale structure
• Non-Gaussian statistics with weak lensing

Jia Liu �https://liuxx479.github.io/, jia.liu@ipmu.jp�Director, Center for Data-Driven Discovery, Kavli IPMU

Associate Professor, Kavli IPMU

Fugaku Projects Overview

Ken Ohsuga

Hideyuki Hotta

Jia Liu

シミュレーションとAIで解き明かす�太陽地球環境変動�Elucidation of the Sun-Earth environment using simulations and AI

領域：④ 基礎科学の発展(②防災・減災、環境問題) �4. Basic Science (2. Disaster prevention and mitigation, Environment)

課題の種類：(c) 標準課題(計算資源のみ)/Numerical resource only

「富岳」成果創出加速プログラム

Program for Promoting Researchers on the Supercomputer Fugaku

課題代表者：堀田　英之

PI: Hideyuki HOTTA

代 表 機 関：名古屋大学宇宙地球環境研究所

Research Institute: Nagoya University/ISEE

​

30

Comprehensive project covering solar interior, surface, corona, interplanetary and geo-magnetosphere

Solar dynamo

Collaboration with Observation

Solar wind

Flare

Coronal heating

Coronal mass ejection

Geo-magnetosphere

Accelerate calc. and collaboration with obs.

Simulation acceleration by AI

Sophisticated collaboration with observation

AIを用いたシミュレーションの抜本的な改革・観測連携の洗練化

Turbulent hierarchy

Difficult to cover in one simulation

Solar atmosphere sim.

RAMENS(Iijima+2017)

Synthesis Obs.

Actual Obs.

Hinode, SOLAR-C、DKIST、Geo-X

Simulation acceleration

Collaboration with Obs.

32

Sub-A

From dynamo in solar interior�to sunspot formation.

太陽内部の乱流磁場生成から�黒点形成までの統一的理解

Sub-B

Solar surface, corona, wind, �and interplanetary space modeling

太陽表面・コロナ・太陽風・惑星間空間�の包括的モデリング

Sub-C�Geo-space modulation by �magnetosphere modeling

超高解像度地球磁気圏モデルによる�ジオスペース変動の解明

Hotta & Toriumi, 2020

Sunspot simulation

R2D2 code

• How are large-scale flow and B constructed?
• How are sunspots created?

Iijima+2023

Solar surface → solar wind at Fugaku

Shiota+2016

Coronal mass�ejection (CME)

RAMENS、SUSANOO codes

• Solar wind acceleration by B
• CME mass/momentum

Matsumoto+2022

Global modeling of �geo-space

Geo-X synthesis

CANS+、Vlasov2D3V codes

• Response of geo-magnetosphere�to the solar wind
• Kinetic effect of plasma

Only R2D2 can deal with �deep CZ to surface in the world

Only RAMES can deal with surface�to solar wind in the world

Hotta & Kusano, 2021

Nature Astronomy

33

８ institute/univ., 24 members

次世代宇宙論サーベイ群のための�多波長宇宙論的シミュレーション

応募する領域：④基礎科学の発展、新領域

応募する課題の種類：(b) 標準課題

​

「富岳」成果創出加速プログラム

課題代表者：LIU Jia

​

代 表 機 関：国立大学法人 東京大学

​

Multi-wavelength Cosmological Simulations for Next-generation Surveys on Fugaku

PI: Jia LIU (CD3, Kavli IPMU)

With contribution from Ken Osato & Hironao Miyatake (Co-PIs)

「富岳」成果創出加速プログラム

Overview of the Project

Multi-wavelength data from

ongoing/upcoming surveys

N-body�Simulations

(1600 runs)

Simulations & Mocks

AI/ML

Emulator

Simulation-�Based

Inference

Fugaku

New Physics

• Dark Matter
• Dark Energy
• Modified Gravity

Overview of the Project

Our Team

PI: Jia Liu

(IPMU)

Co-PI:Ken Osato

(Chiba Univ.)

Co-PI:Hironao Miyatake

(Nagoya Univ.)

Masato Shirasaki

(NAOJ)

Takahiro Nishimichi

(Kyoto Sangyo Univ.)

Satoshi Tanaka

(Kyoto Univ.)

Shogo Ishikawa

(Kyoto Univ.)

Tomotake Matsumura

(IPMU)

Masahiro Takada

(IPMU)

Yici Zhong

(U. Tokyo)

Seongwhan Yoon

(Nagoya Univ.)

Akira Tokiwa

(IPMU)

2022: Machine Learning in Astrophysics

Workshop at KMI, Nagoya University, Nov 21-22, 2022

CENTER FOR DATA-DRIVEN DISCOVERY (CD3)

​

Welcome to collaborate with CD3 researchers every Friday 13:30-17:00 (“Hack Friday”)

​

Mission: advancing our understanding of the universe through data science and AI/ML

​

Website: cd3.ipmu.jp

​

Opening Symposium: April 20, 2023

​

Location: Kavli IPMU, U Tokyo (Kashiwa, Chiba)

​

Stay tuned: (Jan 2024) joint CD3 x CCA workshop in Kavli IPMU on ML x astro

Collaborative Hack

Daily progress report

~ 5 min per group

Add a channel

→ Channel list

# workshop_fugakuai_sep23

Please join the slack channel for Hack

Useful materials

​

Dataset

AstroML (sample notebook)

Solar simulation data (dropbox by Hotta-san)

​

Model

Diffusion model (pytorch)

Diffusion model (JAX)

Groups

Group 1 (Hotta, Iida, Tanimura, Kanbara) Predict n+1 step data from n step data.

​

Group 2 (Harada, Moriwaki, Tokiwa, Inoue) Physics Informed generative AI

​

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

​

Group 4 (Jia Liu, Hanchun Jiang, Ken Ohsuga, Hiroyuki Masaki, Yuta Asahina): Investigate the feasibility of applying transformer on CMB polarization data

Group 1 (Kambara, Iida, Tanimura, Hotta) Predict n+1 step data from n step data.

​

We made the simplest CNN model. It predicts intensity 10 steps later from data of intensity and vertical velosity.

n_param =2, epoch =100

day 2

Tried to use multiple images to predict next image.

-> Memory limit of Google Colaboratory exceed

We use 3 steps to predict 10 steps later, little improvement was seen.

​

​

​

​

2 layer CNN single layer

From density to magnetic field with transformer

Results….

CNN

CNN

CNN

CNN

Attention

True

Pred

Group 2 Physics Informed generative AI

• Target: large-scale structures
• Method: generative AI
• Feature: Physics Informed (power spectrum?)

​

Original github

https://github.com/lucidrains/denoising-diffusion-pytorch

​

Google Colab:

https://colab.research.google.com/drive/1RwOgM_WZyFC63SFMnJsNDqECbdTbAcBi?usp=sharing

​

Reference:

Review on PINN by Karniadakis et al. 2021 https://www.nature.com/articles/s42254-021-00314-5

​

- We use N-body simulation (by Tokiwa)

- Use diffusion model to generate the large-scale structure (left) from random image (right)

- Current problem: the training seems to take long time

Plan for tomorrow:

- train a normal diffusion model and check if the power spectrum is reproduced well (if ok, it's done!)

- train the model with physics sinformation (e.g., loss function = |Ptrue - Prec|)

​

- We tried to generate a snapshot of N-body simulation by diffusion model (data from Tokiwa)

→ Generated figure is somehow almost dark.

We also tried to implement physical information to diffusion model by adding the RMS error of power spectrum to the loss function (under construction)

​

k

P(k)

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Ground Truth

Reconstruction

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Ground Truth

Reconstruction

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Gaussian base

Affine Gaussian

(scaled by x2)

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Gaussian base

Gaussian mixture

Group 3 (Tian, Jingjing, Joaquin, Keiya) Normalizing flow for different geometries

Gaussian base

Gaussian mixture

(3 models)

Group 4: Investigate the feasibility of applying transformer on CMB polarization data

Task 1 (Hiroyuki Masaki, Ken Ohsuga): investigate literature (ADS or arxiv) on applying transformer on imagines, in particular with missing data

Task 2 (Yuta Asahina, Jia Liu): look for notebook with simple transformer architecture that can be used for any data

Task 3 (Hanchun Jiang): prepare CMB polarization data for training

Project plan: Hiroyuki and Ken will review current trend on transformer x image processing (and maybe in astronomy; Hanchun will prepare 1000 small patch CMB polarization maps for training; Finally, apply the network found by Yuta and Jia to 1 map with missing data, then compare the in-filled map (from ML) with the ground truth

Group 4: report for day 1

Data preparation: in progress to generate training data from Planck healpix maps

Vision Transformer vs CNN: transformer may perform better when there is a lot of data, but less environmental friendly (needs a lot of computing power)

​

• “The Universe is worth 64^3 pixels: Convolution Neural Network and Vision Transformers for Cosmology” https://ui.adsabs.harvard.edu/abs/2023arXiv230408192H/abstract

Group 4: report for day 2

Data preparation: 90 training data with map size of 2 degrees and resolution of 0.1arcmin/pix from Planck healpix maps. The shape of each patch is 1200x1200.

We generate the missing data with 100x100 patch randomly. Then we try to reproduce the original data (ground truth) using Vision Transformers.

​

Original Data

Missing Data

Prediction

Cannot predict

Errors?

generate

training

Pathak et al. 2016

Input with gap

Human artist

Deep Learning