NSF AAAC Workshop

Yuan-Sen Ting

The Ohio State University

Report on the Future of AI for the Mathematical and Physical Sciences (AST focused)

About 12 invitees per domain, about 60 in total

List of invitees for Astronomical Science (AST)

Andy Connolly (U Washington - Time Domain)

Stella Offner (UT Austin - ISM / Star Formation - CosmicAI)

Yuan-Sen Ting (Ohio State - Stellar Astrophysics / Galaxy Evolution)

Eric Ford (Penn State - Exoplanet)

Risa Weschler (Stanford - Local Group)

Brice Menard (Johns Hopkins - Science for AI)

List of invitees for Astronomical Science (AST)

Aggelos Katsaggelos (Northwestern - CS - SkAI)

Uros Seljak (Berkeley - Cosmology)

Ann Zabludoff (Arizona - Galaxy Evolution)

Francisco Villaescusa-Navarro (CCA - Cosmology)

Benjamin Wandelt (John Hopkins - Reionization)

Peter Melchior (Princeton - Cosmology)

Salman Habib (Argonne - Cosmology - National Lab - SkAI)

Balancing institutional representation and subdomain expertise

The Need

The upcoming surveys including the Rubin Observatory and Roman Space Telescope will create a data deluge

The Need

Investment in AI remains small compared to the amount already spent on instrumentation

The upcoming surveys including the Rubin Observatory and Roman Space Telescope will create a data deluge

Critical timing for maximizing scientific returns from these
investments

What are some of the main advances so far?

1. More Robust Inference with Simulations

Revolutionizing high-stakes science through novel statistical methods via deep learning

1. More Robust Inference with Simulations

Revolutionizing high-stakes science through novel statistical methods via deep learning

Sharpened constraints in cosmology, gravitational waves, and exoplanet atmospheres, etc.

Simulation-Based Inference transforms the "inverse problem" - using generative models as a surrogate for likelihood / posterior

Bayesian statistical inferences without simplistic human-heuristic summary statistics (e.g., power spectrum)

2. Anomaly Detection

Addressing the astronomical "needle in the haystack" problem

2. Anomaly Detection

E.g.1, Real-time classification of transient events, timely follow-up

E.g.2, Distinction between planetary signals and stellar activity/noise

Addressing the astronomical "needle in the haystack" problem

3. Workflow Optimization

Reduce costs for ground- and space-based telescopes

3. Workflow Optimization

Data processing: e.g., cosmic ray rejection

Image reconstruction and turbulence prediction for adaptive optics

3. Workflow Optimization

Reinforcement learning for instrumentation control
to align multi-mirror telescopes and interferometric arrays

4. Simulation Acceleration

GPU-powered numerical solvers reduce computational requirements

4. Simulation Acceleration

Neural network emulators replace expensive components,
such as sub-resolution physics for multi-scale simulations

Limited by simulation fidelity when modeling multi-scale processes and exposed to unknown systematics

AI for Astronomy - Gaps and Opportunities

Develop benchmark datasets using simulations and real data, implement blind data analysis protocols

Requires serious implementation efforts,
not just conceptual hype

AI for Astronomy - Gaps and Opportunities

Plenty of optimization but few surprising discoveries

Agentic (large language models) systems, and their reasoning capabilities, might offer a high-risk, high-reward pathway

AI for Astronomy - Gaps and Opportunities

Astronomy for AI - Rich Datasets

Minimal ethical concerns or commercial constraints

Safe, low-risk ecosystem, with high-stakes science

Complex, non-proprietary data spanning large dynamic ranges

Astronomy for AI - Rich Datasets

Interesting domain shifts (systematics) between training and test sets, enable study of generalization under realistic constraints

Astronomy for AI - Multimodal Challenges

Combines imaging, spectroscopy, time series, and theoretical models

Astronomy for AI - Multimodal Challenges

Features unique modalities not in mainstream AI research

Astronomy for AI - Multimodal Challenges

Operating in the low signal-to-noise limit and complex/ heterogeneous noise characteristics requires advanced handling

Astronomy for AI - Physics-Informed Learning

Ideal testing ground for incorporating physical laws

Astronomy for AI - Scientific Literature

Exceptionally accessible through arXiv and NASA ADS

Well-curated with long time baselines

Accompanied by publicly available data and code

Agent for science systems for intersectional knowledge domains

Yuan-Sen Ting

Report on the Future of AI for the Mathematical and Physical Sciences (AST focused)

About 12 invitees per domain, about 60 in total

List of invitees for Astronomical Science (AST)

List of invitees for Astronomical Science (AST)

Balancing institutional representation and subdomain expertise

The Need

The upcoming surveys including the Rubin Observatory and Roman Space Telescope will create a data deluge

The Need

Investment in AI remains small compared to the amount already spent on instrumentation

The upcoming surveys including the Rubin Observatory and Roman Space Telescope will create a data deluge

Critical timing for maximizing scientific returns from these investments

What are some of the main advances so far?

1. More Robust Inference with Simulations

Revolutionizing high-stakes science through novel statistical methods via deep learning

1. More Robust Inference with Simulations

Revolutionizing high-stakes science through novel statistical methods via deep learning

Sharpened constraints in cosmology, gravitational waves, and exoplanet atmospheres, etc.

Simulation-Based Inference transforms the "inverse problem" - using generative models as a surrogate for likelihood / posterior

Bayesian statistical inferences without simplistic human-heuristic summary statistics (e.g., power spectrum)

2. Anomaly Detection

Addressing the astronomical "needle in the haystack" problem

2. Anomaly Detection

E.g.1, Real-time classification of transient events, timely follow-up

E.g.2, Distinction between planetary signals and stellar activity/noise

Addressing the astronomical "needle in the haystack" problem

3. Workflow Optimization

Reduce costs for ground- and space-based telescopes

3. Workflow Optimization

Data processing: e.g., cosmic ray rejection

Image reconstruction and turbulence prediction for adaptive optics

3. Workflow Optimization

Reinforcement learning for instrumentation control to align multi-mirror telescopes and interferometric arrays

4. Simulation Acceleration

GPU-powered numerical solvers reduce computational requirements

4. Simulation Acceleration

Neural network emulators replace expensive components, such as sub-resolution physics for multi-scale simulations

Limited by simulation fidelity when modeling multi-scale processes and exposed to unknown systematics

AI for Astronomy - Gaps and Opportunities

AI for Astronomy - Gaps and Opportunities

Develop benchmark datasets using simulations and real data, implement blind data analysis protocols

Requires serious implementation efforts, not just conceptual hype

AI for Astronomy - Gaps and Opportunities

Plenty of optimization but few surprising discoveries

Agentic (large language models) systems, and their reasoning capabilities, might offer a high-risk, high-reward pathway

AI for Astronomy - Gaps and Opportunities

Astronomy for AI - Rich Datasets

Minimal ethical concerns or commercial constraints

Safe, low-risk ecosystem, with high-stakes science

Complex, non-proprietary data spanning large dynamic ranges

Astronomy for AI - Rich Datasets

Interesting domain shifts (systematics) between training and test sets, enable study of generalization under realistic constraints

Astronomy for AI - Multimodal Challenges

Combines imaging, spectroscopy, time series, and theoretical models

Astronomy for AI - Multimodal Challenges

Features unique modalities not in mainstream AI research

Astronomy for AI - Multimodal Challenges

Operating in the low signal-to-noise limit and complex/ heterogeneous noise characteristics requires advanced handling

Astronomy for AI - Physics-Informed Learning

Ideal testing ground for incorporating physical laws

Astronomy for AI - Scientific Literature

Exceptionally accessible through arXiv and NASA ADS

Well-curated with long time baselines

Accompanied by publicly available data and code

Agent for science systems for intersectional knowledge domains

Critical timing for maximizing scientific returns from these
investments

Reinforcement learning for instrumentation control
to align multi-mirror telescopes and interferometric arrays

Neural network emulators replace expensive components,
such as sub-resolution physics for multi-scale simulations

Requires serious implementation efforts,
not just conceptual hype