1

Please contribute one slide to the section for your MPS domain.

​

Navigate to your domain:

​

• Astronomical Sciences (AST)
• Chemistry (CHE)
• Materials Research (DMR)
• Mathematical Sciences (DMS)
• Physics (PHY)

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Astronomical Sciences (AST)

Breakout Discussion

NSF AI+MPS Workshop 2025

2

Feel free to add comments directly in the shared doc:

https://bit.ly/ai-mps-breakout-notes

​

Slides for report: https://bit.ly/ai-mps-breakout-report

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Your Name, Affiliation (MPS Domain)

Key Issue for Your Domain

• Describe here one key issue for your domain that you want to make sure is discussed and/or included in the white paper.
• What is the primary takeaway? What questions do you have for the group?

​

Domain-Specific Case Study/Example

• Describe here one case study or example from your experience of a success story from your domain related to AI.
• This could be a key result in research, strategies for incorporating AI, building interdisciplinary collaborations, etc.

3

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Brief description of how you’ve used AI in your work.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Cecilia Garraffo, CfA Harvard & Smithsonian (AST)

Key Issue for Your Domain

• Simulations are a huge bottleneck for astrophysics. We need to combine AI frameworks that guarantee to respect physical laws (e. g. symmetries) with scalable generative models, ensuring both local and global conservation laws. We need them to be flexible enough to discover new physics / symmetries.
• How can we standardize and benchmark AI approaches that integrate global physical constraints? Which emerging architectures or mathematical tools can handle multi-scale, long-range dependencies effectively? What interdisciplinary training strategies best prepare a workforce to advance physical GenAI?

Domain-Specific Case Study/Example

• Denoising Hamiltonian Networks (DHN) for chaotic systems (e.g., double pendulum): These demonstrate improved long-term accuracy, reduced error accumulation, and robust parameter inference by enforcing global physical constraints. How do we scale these to complex physical systems?

4

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Photo of you

I am the director of AstroAI, an institute to conducts AI driven research in astrophysics. I focus on physical, probabilistic models, representation learning, multi-modal models for astronomy, and physical generative AI.

​

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Andrew Connolly, University of Washington (AST)

Key Issue for Your Domain

• Effective integration of AI/ML with a new generation of surveys requires more than just the adoption of current AI methodologies. It requires an ecosystem that can manage data access, analysis frameworks, and computational resources together with support for the development of science focused AI methodologies.
• How do we understand the difference/gaps between industry funded tools and the needs for science to ensure the development of appropriate methods (e.g. integration of physics constraints)

Domain-Specific Case Study/Example

• Integration of ML/AI applications in the optimization of the performance of telescopes (scheduling observations, optimizing image performance, and removing spurious data) has significantly increased the efficiency of astronomy surveys. The is significant opportunity to build on these individual approaches when designing new experiments

5

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I develop AI/ML and software for working with large data sets from Astronomical Observatories. These include AI methods to optimize observatory performance and large scale machine learning frameworks

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Eric Ford, Penn State (AST)

Key Issue for Your Domain

• AI is on track to create a new crisis of confidence in AST and science in general.
• AI allows researchers to generate low-quality results faster.
• AST prioritizes novelty and speed over quality and thoughtful application, encouraging researchers to dilute scientific literature with low-quality “results”.
• Checking/explaining AI’s mistakes takes more time and expertise than generating new results, causing many claims to go unchecked.
• We are on a trajectory to dilute the scientific literature with AI-accelerated “results” that cause scientists and the public to lose trust in the scientific enterprise.

Domain-Specific Case Study/Example

• We’re developing physics-informed ML models for extracting the Doppler signal of low-mass planets from spectroscopic time-series corrupted by stellar variability.
• Incorporating domain knowledge (“feature engineering”, symmetries, etc.) can allow even very simple models to provide factor of ~2-3x benefits over conventional AI, plus side benefits of faster training and greater explainability.
• Unfortunately, this approach takes more time of human experts to merge knowledge from astronomy, electronics, mathematics, statistics, and AI/ML.

6

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Developing physics-informed ML models for analyzing Doppler exoplanet surveys data to overcome the barrier of intrinsic stellar variability and detect potential Earth-analog planets.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Salman Habib (PHY/AST)

Key Issue for Your Domain

• I think the key issue is to develop methods that are demonstrably better than the SOA when applied to actual problems – not toy problems and not method demo exercises. The problems range from very specific instances to very complex situations involving major collaborations. Often people prefer simpler, robust techniques to fragile methods. We need to take AI-enhanced approaches forward so that they become controlled tools; we are not there yet.
• My key question is how to identify major areas where progress of the kind mentioned above is possible, demonstrating a true and seriously significant advantage.

Domain-Specific Case Study/Example

• One example from cosmology is emulation of summary statistics and realistic predictions for observations (based on large-scale simulation-based computational pipelines) with applications to simulation based inference and survey optimization.

7

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Large-scale surrogate models based on extreme-scale HPC codes, domain-specific foundation models, applications of LRMs to code modernization and translation – long history of collaboration with computer scientists and statisticians

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Peter Melchior, Princeton (AST)

Key Issue for Your Domain

• Black-box ML/AI approaches are antithetical to most analysis tasks in the discipline: We want more than predictions, we want insights. This, and the lack of understanding of current methods, creates significant friction and resistance.
• We either have to develop our own methods for AI in AST or restrict its use to the engineering side of the discipline (calibrations, telescope control, etc.)

Domain-Specific Case Study/Example

• Most interesting and effective result(s) are from ML-aided discovery, where too much or too complicated data need to parsed for too subtle an effect. The goal is present meaningful, actionable summaries: E.g. unifying microlensing degeneracies (theory), finding outliers, linking different wavelength regimes (observations), optimizing target selection (engineering)

8

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I develop ML (diffusion models, flows, auto- encoders) with physics-preserving components to extract highly informative representation from observations. The goals are high-precision galaxy models and data-driven discovery

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Brice Ménard, Johns Hopkins Univ (AST)

Key Issue for Your Domain

• AI has accelerated/optimized many things in astrophysics but, so far, has not clearly led to new insights or better physical understanding. Will this change?
• There is, in principle, a lot to learn from neural encodings but the black box reputation of neural networks appears to discourage most people from investigating them.
• Can we use AI to analyze the astro literature and gain new insights?

​

Domain-Specific Case Study/Example

• success stories:
• Simulation-based inference
• manipulating more complex shapes in signals (galaxy morphologies) or noise (PFS modeling for exoplanet detection)

9

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Current research: physics of learning in artificial and biological systems.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Stella Offner, UT Austin (AST)

Key Issue for Your Domain

• A recent survey carried out by the American Astronomical Society with 500 respondents found a number of widespread concerns that pose significant barriers to community adoption of ML/AI in astronomy research:
• Skepticism, lack of trust in AI/ML results
• Concerns about the impact of (generative) AI on the environment
• Lack of training in AI/ML methods
• How do we effectively train the community in ML/AI methods, increase community confidence, and increase engagement/participation?

Domain-Specific Case Study/Example

• Data segmentation of spectral line cubes, e.g., to identify features of young stars interacting with their environments. These studies use multi-physics simulation data to build training sets to identify complex features in noisy, high-d data.
• In general, there has been a significant success in using simulations to build high-quality training sets (e.g., simulation-based inference as used in cosmology).

​

10

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Director of CosmicAI.

My star formation research has used both supervised ML for classification, data segmentation, system modeling (CNNs, diffusion models, neural operators) and unsupervised ML for classification, discovery (UMAP, SOM) in big data

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Yuan-Sen Ting, The Ohio State University (AST)

Key Issue for Your Domain

• While there is strong interest in applying AI techniques to astronomical research, there's a lack of initiative and opportunity for researchers who are pushing the frontier of AI methods specifically tailored for astronomy.�
• This situation is analogous to what numerical theorists faced 10-20 years ago — astronomy departments considered them not 'astronomy enough,' while mathematics departments viewed them as not 'math enough.' Most departments just want researchers who can apply AI, not develop it.

Domain-Specific Case Study/Example

• Agentic AI for analyzing James Webb telescope spectra has started to gain traction — researchers have successfully adopted AI agents as solvers to analyze out-of-domain outliers with promising results.

​

11

NSF AI+MPS Workshop: March 24–26, 2025; MIT

My research focuses on two domains: using AI as a surrogate for high-dimensional Bayesian inference ��and deploying AI agents to autonomously solve astronomical problems.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Benjamin Wandelt, Johns Hopkins University (AST)

Key Issue for Your Domain

• Integrating AI into mathematical and physical sciences demands robust validation frameworks that establish trust beyond a small group of experts. A major challenge is moving from proof-of-principle demonstrations (e.g., on simulations) to actual applications on data. This requires significant development resources and expertise to verify reliability, robustness, reproducibility, and alignment with physical principles. Success requires interdisciplinary collaboration, standardized scientific benchmarks, and FAIR principles enabling proper scrutiny of both data and models—ensuring AI enhances rather than undermines scientific rigor.
• We have the potential to make current methods seem quaint and antiquated in 5-10 years. How do we get there?

Domain-Specific Case Study/Example

• AI-accelerated modeling allows costly physical models to scale to new regimes; this has the potential to bring previously separate communities (e.g. cosmology and galaxy formation) together. Great opportunities for interdisciplinary work!

12

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I develop AI/ML methods to solve computational astrostatistics problems for cosmological surveys and generative models to accelerate non-linear simulations in astrophysics and cosmology.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Ann Zabludoff, University of Arizona (AST)

Key Issue for Your Domain

• NAS defined two AI challenges for science: advanced literature search and complex hypothesis generation (AI for Scientific Discovery, NAS Workshop, Oct 2023)
• How do we use AI to reliably leapfrog the painstaking work required to create an original plot from the literature, track the progression of key measurements and concepts within or across disciplines, tie disparate findings into a coherent “worldview” that generates new hypotheses? And how can the most impactful new hypotheses be selected, ie, what methods improve AI’s “scientific taste”?

Domain-Specific Case Study/Example

• NASA Astrophysics Data System (ADS) now employs AI for semantic searches, chatbot assistance, graph-based citation data, building custom search tools
• How do we move beyond lists of paper links?

13

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI-based literature search and visualization to instantly return data in tabular, graphical, and/or textual form and to which causal AI methods can be applied for automated hypothesis generation. Incorporation of user feedback into model.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Francisco Villaescusa-Navarro, Flatiron Institute (AST)

Key Issue for Your Domain

• The most sophisticated simulations to-date are just (crude) approximate representations of the real Universe. Different simulations tend to give vastly different predictions. How can ML help in this regard?
• We need a collective effort to move decades-old fortran/c code to run into modern GPU-based architectures.

Domain-Specific Case Study/Example

• In general, training any ML model (CNN/GNN/transformer) on a particular hydrodynamic simulation code will fail when testing on another simulation code. We found that using positions and velocities works! ML told us that phase space is robust to uncertain subgrid physics, and we derive equations about it.
• Building projects like CAMELS and DREAMS requires experts on numerical simulations, galaxy formation, cosmology, particle physics, and machine learning.

14

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I run very large suites of state-of-the-art cosmological simulations and use machine learning to extract hidden patterns, marginalize over baryonic effects…etc. Examples: Quijote, CAMELS, DREAMS, Backlight. Over 100k sims and 4Pb of data

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Uros Seljak, UC Berkeley/LBNL (AST)

Key Issue for Your Domain

• How do we handle model misspecification in Simulation Based Inference given the lack of ab initio model (e.g. galaxy formation in cosmology surveys)?
• How do we search for and/or quantify unknown systematics in the data?

​

Domain-Specific Case Study/Example

• AI for Cosmology: outlier/anomaly detection, generative modeling (fast simulations) and posterior analysis are the main desiderata of AI for cosmology. They can all be obtained using a single architecture (Normalizing Flows). Robustness against the model misspecification can be tested using consistency across multiple scales. Multi-Scale Flow achieves all these goals in a single package.

15

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Development of novel AI tools for astronomy and cosmology data analysis

​

Development of physics inspired tools for AI.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Aggelos Katsaggelos, Northwestern U (ECE/SkAI)

Key Issue for Your Domain

• Lack of interpretability of the results by neural networks and lack of uncertainty quantification makes them not trustworthy and scientifically useful.
• Very often, while numerous AI approaches are developed and proof of concept results are presented, they are not tested or optimized in a full science pipeline, i.e., working fast, efficiently, and robustly at scale.
• The low barrier to entry in the application of neural networks results in many results for which the appropriate tool was not used or was not used correctly.

Domain-Specific Case Study/Example

• Simulation of single and binary star evolution using PINNs and track and profile interpolation.
• Development of models for describing the interaction between computer science and citizen science for glitch classification (Gravity Spy project).

16

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Development of ML/DL approaches with diverse applications, ranging from astronomy to medicine and cultural heritage. Some recent results deal with self-supervised, multiple instance, and hybrid learning.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Risa Wechsler, Stanford/SLAC (AST)

Key Issue for Your Domain

• How do we fully leverage the unique aspects of astronomy [large and complex open data, open-source code, physics-based models, well-defined literature linked to data, broad public interest] for both discovery and methodology development?
• How do we meet this AI moment thoughtfully and responsively in a way that brings the whole community along AND improves and accelerates rigorous high-quality work, physical understanding, and discovery?

Domain-Specific Case Study/Example

• The astronomical data this decade will be exciting and complex, requires new techniques to fully leverage. How do we build fast, flexible models that are both physics based and data driven, to enable large-dimensional inference from multi-model data? Challenges of model misspecification and need for robust uncertainty quantification. Example with progress: fast simulation-based inference for strong lensing.
• Our new center has had some early success in sparking collaborations across expertise and domains, including on 3D inference from data and AI agents for research.

17

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Director of KIPAC and new Center for Decoding the Universe @ Stanford.

​

AI at the interface of modeling and large cosmological surveys, incl fast modeling of complex problems,high-dimensional inference, classification of large data sets.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI + AST Questions to Consider

18

• Besides making things "faster" / better inverse problem solving, can AI make physically motivated / interesting discoveries?�
• Can AI make interesting discoveries by itself? If not, to what extent can it contribute?�
• What is the utility of AI in underexplored subdomains of astrophysics?�
• What are the key challenges in creating a healthy ecosystem of AI × Astronomy?

Driving Question: How can the MPS domains best�capitalize on, and contribute to, the future of AI?

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Chemistry (CHE)

Breakout Discussion

NSF AI+MPS Workshop 2025

19

Feel free to add comments directly in the shared doc:

https://bit.ly/ai-mps-breakout-notes

​

Slides for report: https://bit.ly/ai-mps-breakout-report

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Suri Vaikuntanathan , UChicago (DMR)

Key Issue for Your Domain

• Machine learning guided by physical laws/statistical mechanics/biological examples
• Can AI guided by physical laws be more “generative” ?

​

Domain-Specific Case Study/Example

• Reimagining generative diffusion process using physical noise sources (https://arxiv.org/abs/2411.07233)
• Physical noise sources incorporated into diffusion models can improve generative properties. .

20

NSF AI+MPS Workshop: March 24–26, 2025; MIT

​

Physics inspired Reinforcement learning, generative diffusion, associative memory for sequence recall, protein design.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Olaf Wiest, University of Notre Dame, C-CAS (CHE)

Key Issue for Your Domain

• The availability and quality of training data in (synthetic) chemistry limits the ability to construct models that can navigate the high-dimensional space for optimization of properties, yields, selectivity etc. More relevant and higher quality data only partially address this problem in an essentially infinite chemical space; more data-efficient methods that incorporate chemical principles are needed.
• What are the right data and quality controls ? How do we make this data available to the community ? Who owns the data and the models build on it ?

Domain-Specific Case Study/Example

• The application of Bayesian methods to systematic and data-efficient reaction optimization (Shields et al., Nature 2021) rapidly became the State-of-the-Art in academia and industry.
• Using an easy-to-use platform to address a key problem in synthetic organic chemistry was the key to widespread adoption across many use cases.

21

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I direct the NSF Center for Computer Assisted Synthesis (C-CAS) which aims to change the synthetic chemistry from an intuition- to a data-driven science.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Pratyush Tiwary, U Maryland (CHE)

Key Issue for Your Domain

• In computational chemistry, we often lack high-quality data at the timescales and system sizes required to capture rare events (e.g., slow protein conformations or RNA folding). This shortage of “right data” undermines many AI-driven methods, which then struggle with out-of-distribution generalization or extrapolation beyond the training set.
• How do we ensure AI predictions remain physically valid in data-sparse settings?
• What methods promote robust out-of-distribution performance and true extrapolation?
• Chemistry needs precise environment. Which strategies best incorporate environmental factors (e.g., temperature, pH) into AI models for more realistic outcomes?

Domain-Specific Case Study/Example

• A recent success (Herron et al., PNAS 2024) predicted phase transitions in Ising and RNA systems without sampling near the critical point. By embedding thermodynamics into AI, we inferred critical exponents from sparse data—showcasing statistical physics-informed AI for emergent phenomena.

22

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I merge AI with statistical physics to simulate protein, crystals & RNA across otherwise unreachable timescales and with limited training data.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

23

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Sijia Dong, Northeastern University (CHE)

Key Issue for Your Domain

• There are not enough highly accurate electronic structure data (e.g., those for electronic excited states) for big-data research due to the high computational cost of generating such data.
• We need to generate such data. We may also need to design our projects differently. Will database-building work be recognized as being valuable? How should such database be organized? Should we establish a standard for it? And what will such standard look like?

Domain-Specific Case Study/Example

• In our work (Dong, et al. Chem. Sci. 2021, Jayee, et al. in preparation), we modified the problem to make use of an intermediate property so that we do not need to generate all the structure-property data and the AI model can be readily transferable to different structures for electronic-excited-state property prediction.

24

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I use AI to accelerate or replace expensive quantum mechanics based simulations (in both classical and quantum computing) of large molecular systems and materials and to design macromolecular photocatalysts.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Heng Ji, University of Illinois Urbana-Champaign,MMLI (CS)

Key Issue for Your Domain

• How to develop science inspired AI techniques?
• How can we effectively represent molecules for AI algorithms?
• How to equip large language models with domain knowledge for scientific discovery?
• How to enable agent and human scientist collaboration for scientific experiment planning?

Domain-Specific Case Study/Example

• Joint natural language and molecule representation for drug discovery
• Combining large language model’s parametric knowledge and knowledge graphs for OPV material discovery

25

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I’m the lead of AI thrust for AI Institute on Molecule Synthesis (MMLI), Founding Director of Amazon-UIUC AI Center and CapitalOne-UIUC AI Center

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Shuwen Yue, Cornell (CHE)

Key Issue for Your Domain

• Physical interpretability in model training – How can we guide model development to predict the “right answers for the right reasons.” This involves physical constraints and evaluation metrics on multiple scales from electronic to atomistic to collective properties
• Frameworks for validation and benchmarking – how do we develop reliable assessment metrics beyond simple error metrics to ensure models generalize well and perform correctly in real applications

​

Domain-Specific Case Study/Example

• Development and application of ML potential for thermodynamic quantities
• Our work on developing cheaper and efficient uncertainty quantification schemes for chemical and materials properties

26

NSF AI+MPS Workshop: March 24–26, 2025; MIT

ML potential development and application for liquids/interfacial thermodynamic quantities, uncertainty quantification, materials design

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Huimin Zhao, University of Illinois Urbana-Champaign (CHE)

Key Issue for Your Domain

• How can we predict enzyme properties (e.g. activity, specificity, selectivity, stability, solubility) from protein sequences?
• How to generate enzymes with new-to-nature functions?

​

Domain-Specific Case Study/Example

• Developed an AI tool called CLEAN to predict enzyme function from protein sequence (Science 2023)
• Developed an AI-powered autonomous protein engineering platform (Nature Communications in revision)

27

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I’m the Director of NSF AI Institute for Molecule Synthesis (MMLI) and NSF iBiofoundry

​

My research program focuses at the interface of synthetic biology, AI/ML, and laboratory automation.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Grant Rotskoff, Stanford (CHE)

Key Issue for Your Domain

• Methods for integrating multimodal experimental data are not widely used and multi-objective optimization under-represented in benchmarks.
• How do we robustly constrain model outputs using “medium throughput” experimental data?
• Are priors for chemical machine learning based on large scale pre-training empirically useful?

​

Domain-Specific Case Study/Example

• Pre–trained molecular transformers accurately predict structure from conventional 1D NMR data with modest amounts of simulated spectral data: https://pubs.acs.org/doi/full/10.1021/acscentsci.4c01132.

28

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Solving high-dimensional PDEs arising in physics, variational inference with generative models, large-scale pre-training for efficient latent representations.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Brett Savoie, University of Notre Dame (ChE)

Key Issue for Your Domain

• Open-source all federal funded fundamental research content. For historical reasons the research records of the past century are largely locked behind paywalls in semi-digital form that leaves them unused.
• The absence of mid-scale (10-20M$) data generation efforts. Chemistry and materials have a bunch of shovel ready but only incompletely executed data generation projects that would cost relatively little but would be broadly very useful:

(i) Organic synthetic reaction chemistry: yield and side-product information across informative substrate scopes, preparation scales, with systematic protocols.

(ii) Materials degradation data: stressor-specific aging, degradation timescales, degrants, and mechanical failure data across material classes.

(iii) Molecular thermodynamics: standard energies, phase diagrams, solubilities, heat capacities, thermal stabilities, densities, etc.

• Cultural changes: There are several cultural dislocations on the horizon:

(i) The meaning of expertise: expertise has historically only defined as something inside a human brain. That is already changing.

(ii) Derivative work will evaporate: Taking “X from field Y and applying it to field Z” is something that AI can already do very effectively.

(iii) “Shut up and Calculate!”: The Mermin quote could start applying to most of the discovery process.

​

29

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Our group develops new generative architectures for structure prediction and chemical design problems. We also work on ML/Physics-based methods for predicting chemical reaction networks.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Johannes Hachmann, U Buffalo (CHE)

Key Issue for Your Domain

• Making AI/ML work in the small data scenarios we typically encounter in chemistry.
• We have to develop methods that can make do with modest/heterogeneous/ heteroscedastic data sets. At the same time, the funding agencies need to take over the hosting of data in similar fashion to them providing leadership computing facilities.

Domain-Specific Case Study/Example

• My group developed a physics-infused DNN model to predict the index of refraction of polymers that legerages knowledge of the underlying Lorentz-Lorenz eqn.
• We’ve been developing the ChemML package that is geared towards making ML accessible to the community.

30

NSF AI+MPS Workshop: March 24–26, 2025; MIT

My group uses AI/ML in chemical and materials discovery and design.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI + CHE Questions to Consider

31

• What are flagship examples of AI+CHE that will help demonstrate where the field is now?
• Are the following subdomains the best way to categorize Chemistry?
• Drug Discovery
• Catalysis and Reaction Engineering
• Materials Chemistry
• Quantum Chemistry and Molecular Simulations
• Cheminformatics & Structural Analysis
• What are the most important priorities and opportunities in AI+CHE in the next 5 years?
• What are the key challenges to address in AI+CHE?
• What does CHE have in common with the other MPS disciplines when it comes to AI?

Driving Question: How can the MPS domains best�capitalize on, and contribute to, the future of AI?

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Materials Research (DMR)

Breakout Discussion

NSF AI+MPS Workshop 2025

32

Feel free to add comments directly in the shared doc:

https://bit.ly/ai-mps-breakout-notes

​

Slides for report: https://bit.ly/ai-mps-breakout-report

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Your Name, Affiliation (MPS Domain)

Key Issue for Your Domain

• Describe here one key issue for your domain that you want to make sure is discussed and/or included in the white paper.
• What is the primary takeaway? What questions do you have for the group?

​

Domain-Specific Case Study/Example

• Describe here one case study or example from your experience of a success story from your domain related to AI.
• This could be a key result in research, strategies for incorporating AI, building interdisciplinary collaborations, etc.

33

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Brief description of how you’ve used AI in your work.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Andrew Ferguson, UChicago (DMR + CHE)

Key Issue for Your Domain

• General strategies for incorporating hard constraints or informative physical priors into deep generative models
• Sustainability and energy cost of model training, efficient learning strategies (e.g., adapters, LoRA)

Domain-Specific Case Study/Example

• Multi-modal learning of a joint latent space of protein annotations and sequences to enable a “ChatGPT” for proteins with experimental wet lab validations�“Natural Language Prompts Guide the Design of Novel Functional Protein Sequences” https://doi.org/10.1101/2024.11.11.622734
• Democratization of deep generative protein design, compositionality of different known facets protein function, scope for supernatural function via active learning?

34

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI/ML for dimensionality reduction, enhanced sampling methods, collective variable discovery, active learning, rare event dynamics, active learning materials and molecular discovery, deep generative protein design

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Yaroslava Yingling, NC State University (DMR+CHE)

Key Issue for Your Domain

• Developing guardrails for materials data is essential to ensure accuracy and reliability as the volume of publications and datasets continues to grow. AI currently doesn't differentiate between non-experts and expert opinions. To improve AI’s reliability - we need chemistry and physics-based guardrails. Knowledge weighting is essential to assign greater importance to seminal sources over less-reliable content.
• We need data validation methods to help identify the most reliable data for AI-driven materials research.

Domain-Specific Case Study/Example

• By leveraging advanced data imputation and augmentation methods, we can overcome the limitations of small experimental datasets and effectively train supervised machine learning models to optimize experimental synthesis.
• “Machine Learning and Small Data-Guided Optimization of Silica Shell Morphology on Gold Nanorods” https://pubs.acs.org/doi/full/10.1021/acs.chemmater.3c03204
• Developing knowledge hub by integrating GenAI with knowledge graphs that enhances data integration and analytics, fostering convergent interdisciplinary research

35

NSF AI+MPS Workshop: March 24–26, 2025; MIT

uncertainty quantification, small data, data imputation, data integration via knowledge graphs, data fusion, inverse materials design, machine learning force fields, GenAI research workflows

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Camille Bilodeau, University of Virginia (DMR + CHE)

Key Issue for Your Domain

• For most molecular design problems, there is not enough data to build a robust, generalizable predictive model and while one can always generate more data, we also need to find ways to do more with the data we can reasonably obtain
• Possible solutions: Physically motivated constraints/priors, multi-modal/multi-fidelity training, methods for evaluating generalizability and uncertainty, community emphasis on sharing and cultivation on new, quality-controlled datasets

Domain-Specific Case Study/Example

• Our group endeavors to build robust, high throughput deep learning models that can be used in a molecular design context by leveraging 1) knowledge of biophysics, and 2) knowledge of sources of error to guide and constrain model development
• We’ve developed a hierarchical graph convolutional network to represent peptides with arbitrarily complex synthetic features and are exploring multi-fidelity approaches for learning efficiently from heterogeneous datasets

36

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Graph-based learning for predicting molecular properties, generative models for molecular discovery, multi-fidelity learning using knowledge of simulation/experimental errors, baking classical and statistical thermo into model design.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Rebecca (Becky) Lindsey, U. Michigan (DMR +CHE)

Key Issue for Your Domain

• Effective student training - ML/AI is a fact of life. Most students come in with some “grass-roots” experience with these tools, but no formal training on how they work or how to use them; most faculty are ill-equipped to address this in the classroom. Additionally, increased emphasis need to be placed on the intersection of ML/AI and physical sciences - students are increasingly exhibiting an “apply ML first” thinking pattern, failing to ask “does a physics-based solution already exist for this problem?
• Support for developing integrated ML/AI courses; centralized teaching materials; faculty education opportunities.

Domain-Specific Case Study/Example

• I have developed a computational modeling course that takes students from fundamental physical principles to machine-learning accelerated methods while making use of generative AI. The course teaches students strengths and weaknesses and responsible use of both physics- and ML/AI- based approaches and how powerful they can be when combined.

37

NSF AI+MPS Workshop: March 24–26, 2025; MIT

ML interatomic models (IAM) and strategies for uncertainty quantification, active, learning, etc.; application to atomististic modeling and optimization of non-equilibrium phenomena relevant to synthesis science; education in computation DMR+CHE

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Chad Risko, University of Kentucky (DMR + CHE)

Key Issue for Your Domain

• Materials are often heterogeneous or disordered, which means that data from experiment or computation are likewise heterogeneous / disordered. However, many materials systems are treated as being pristine / ordered, meaning that we may lose important physicochemical characteristics that are key to developing models.
• How do we handle heterogeneity / disorder? What physical principles can be readily implemented into the models? How can uncertainty quantification be implemented? How do we make model training & data access sustainable and open?

Domain-Specific Case Study/Example

• We developed ML models to predict electronic, redox, and optical properties of medium-to-large sized π-conjugated molecules, and reported the hierarchy of approaches used to derive the models (DOI: 10.1039/D2SC04676H)
• The data and ML models were developed with input from synthetic chemists, physical chemists, and computational mechanists, and made available in easy-to-use formats on OCELOT

38

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Data infrastructure development, including considerations of pluralistic vs universal ontologies; creating open-access data and ML models; automated to autonomous experiments and data collection / reproducibility

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

39

NSF AI+MPS Workshop: March 24–26, 2025; MIT

40

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Keith A. Brown, Boston University (DMR + CBET)

Key Issue for Your Domain

• Fusing experimental and computational data. Specifically that experimental data is generally low throughput but closer to reality (with appropriate caveats about uncertainty) while computation is generally higher throughput but could have a larger disconnect with reality
• We need better transfer learning strategies that connect data across scales and ideally employ physics principles to make this more efficient. A key feature of this would be using experiment in a targeted fashion to optimize or improve computation

Domain-Specific Case Study/Example

• Self-driving labs are generally a success story in terms of doing research faster and more effectively. In my own research, we have performed a multi-year campaign that allow us to uncover world-record mechanical performance.

41

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Autonomous experimentation in which AI is used to select experiments that are performed by robotic systems with a focus on polymer materials

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Suri Vaikuntanathan , UChicago (DMR)

Key Issue for Your Domain

• Machine learning guided by physical laws/statistical mechanics/biological examples
• Can AI guided by physical laws be more “generative” ?

​

Domain-Specific Case Study/Example

• Reimagining generative diffusion process using physical noise sources (https://arxiv.org/abs/2411.07233)
• Physical noise sources incorporated into diffusion models can improve generative properties. .

42

NSF AI+MPS Workshop: March 24–26, 2025; MIT

​

Physics inspired Reinforcement learning, generative diffusion, associative memory for sequence recall, protein design.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Your Name, Affiliation (MPS Domain)

Key Issue for Your Domain

• Describe here one key issue for your domain that you want to make sure is discussed and/or included in the white paper.
• What is the primary takeaway? What questions do you have for the group?

​

Domain-Specific Case Study/Example

• Describe here one case study or example from your experience of a success story from your domain related to AI.
• This could be a key result in research, strategies for incorporating AI, building interdisciplinary collaborations, etc.

43

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Brief description of how you’ve used AI in your work.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI + DMR Questions to Consider

44

• …

• What are flagship examples of AI+DMR that will help demonstrate where the field is now?
• What key subdomains are poised to make best use of AI in Materials Research?
• Soft Materials – polymers, gels, rubbers, networks
• Biomolecular Materials – proteins, peptides, nucleic acids
• Hard Materials – catalysts, electronic and optical materials
• Biomedical Materials – small molecule ligands and drugs, adjuvants, vaccines
• Quantum Materials – substrates for qubits, quantum sensing and information
• Automated Robotics & Self-Driving Labs
• What are the most important priorities and opportunities in AI+DMR in the next 5 years?
• What are the key challenges to address in AI+DMR?
• What does DMR have in common with the other MPS disciplines when it comes to AI?

Driving Question: How can the MPS domains best�capitalize on, and contribute to, the future of AI?

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Mathematical Sciences (DMS)

Breakout Discussion

NSF AI+MPS Workshop 2025

45

Feel free to add comments directly in the shared doc:

https://bit.ly/ai-mps-breakout-notes

​

Slides for report: https://bit.ly/ai-mps-breakout-report

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Your Name, Affiliation (MPS Domain)

Key Issue for Your Domain

• Describe here one key issue for your domain that you want to make sure is discussed and/or included in the white paper.
• What is the primary takeaway? What questions do you have for the group?

​

Domain-Specific Case Study/Example

• Describe here one case study or example from your experience of a success story from your domain related to AI.
• This could be a key result in research, strategies for incorporating AI, building interdisciplinary collaborations, etc.

46

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Brief description of how you’ve used AI in your work.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Soledad Villar, Johns Hopkins (DMS, computational math, modl)

Key Issue for Your Domain

• Mathematical research in AI:
• Explain the behavior of AI models (e.g. training dynamics, generalization properties of trained models, sample complexity, etc).
• Use mathematical principles to design machine learning models (e.g. group equivariant models implemented via representation theory or invariant theory).
• Use of AI to prove theorems, verify proofs, or make conjectures.

Domain-Specific Case Study/Example

• My work uses techniques from algebra (invariant theory, galois theory, representation theory) to design machine learning models that are (approximately) invariant/equivariant with respect to group actions.
• Examples: machine learning on point clouds (applications to cosmology/computer vision), graph neural networks, equivariant self-supervised learning.

47

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Equivariant machine learning, graph neural networks, mathematical theory of deep learning

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Jeremy Kepner, MIT LLSC (DMS)

Key Issue for Your Domain

• There is a significant need to develop detailed and reproducible measurements of AI systems that can be the target for theoretical analysis.
• Current approaches rely on ad-hoc measurement of operational AI systems that are a difficult target for theoretical investigations.

Domain-Specific Case Study/Example

• Among the phenomena that have been observed from operational AI systems a few seem to suggest possible paths for the theoretical inquiry. Among these are the surprising:
• Effectiveness of pruning and quantization
• Ability of l-2 (least squared error) cost functions to produce high Top N scores
• Weak dependence on the shape of the non-linear response function
• Oscillation between arithmetic semiring (weights) and tropical semiring (biases)
• Ubiquity of heavy-tail distributions in training data

48

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I lead the Lincoln Laboratory Supercomputing Center. We enable thousands of AI researchers at MIT. Our interest is in developing more predictable approaches to AI development.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

René Vidal, University of Pennsylvania (DMS/CISE)

Key Issue for Your Domain

• In deep learning, there is a big gap between theory and practice. I conjecture that some of the practical success is due to the hierarchical and compositional structure of the data. However, most theoretical results are worst case, and do not sufficiently exploit the structure of the data.
• How do we develop new deep learning theory that exploits data structure? For example, what data structures can be learn by a diffusion model?

Domain-Specific Case Study/Example

• My work focuses on understanding the learning dynamics of overparametrized networks, showing how certain phenomenon (robustness, neural collapse) are related to data structure. For example, we have shown that a robust classifier exists if and only if the data from each class localizes on regions of small volume.

49

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Deep Learning Theory: non-convex optimization, overparametrization, learning dynamics

​

Trustworthy AI: interpretability, robustness.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Jeremy Avigad, Carnegie Mellon (DMS)

Key Issue for Your Domain

• A key challenge is to make AI tools and capabilities accessible to the mainstream mathematical community.
• Tools need to be adapted to mathematical research needs; we need collaboration between AI developers and mathematics researchers; we need training, documentation, exposition, and technical support for the mathematics community (and these need to be incentivized somehow).

Domain-Specific Case Study/Example

• I am among the organizers of a Simons Institute for the Theory of Computing and SLMath Joint Workshop, AI for Mathematics and Theoretical Computer Science.
• Response was so strong we have had to close registration (even without funding).
• The workshop will only “break the ice” for mathematical researchers.

50

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Interactive theorem proving, automated reasoning, neuro-symbolic methods for mathematics.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Robert Ghrist, Penn (DMS/ENG)

​

Key Issue for Your Domain

• How to facilitate researchers using generative AI tools to augment/expand research capabilities beyond currently imagined use-cases?
• In Mathematics, research is often judged by the inf-norm (you’re as good as your worst result); this leads to mis-valuing what AIs can do in research
• How to best experiment & judge experimentation while avoiding backlash?�

Domain-Specific Case Study/Example

• Case studies: using generative AI tools / LLMs in research
• DeepSearch for literature search / interdisciplinary connections
• Group plans (OpenAI/Claude Teams) to conduct collaborative research
• Exploring the latent space of proofs with different models

51

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Research in : distributed systems, topological data analysis, sheaves for neural nets, algebraic Laplacians, lattice valued networks��Generative-AI-led research

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Lars Ruthotto, Emory University (DMS, computational math)

Key Issue for Your Domain

• Mathematical analysis of complex AI systems
• Scientific Computing for Machine Learning
• Continuous-time probabilistic models and links to optimal control
• Mixed-precision algorithms for machine learning
• Improving efficiency of architectures / training
• Machine Learning for Scientific Computing
• Approximating solutions of high-dimensional PDEs
• Develop generative AI models for high-dimensional probability
• Learnable numerical algorithms
• Provable convergence / sample-efficiency / error estimates / reproducibility

Domain-Specific Case Study/Example

• The frontier of simulation-based inference, Cranmer, Brehmer, Louppe, PNAS 2020
• Algorithms for Solving High Dimensional PDEs: From Nonlinear Monte Carlo to Machine Learning, E, Han, Jentzen, Nonlinearity 2022

52

NSF AI+MPS Workshop: March 24–26, 2025; MIT

computational mathematics ⇔ machine learning

​

Applications in inverse problems, data assimilation, optimal control

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Sergei Gukov, AIM & Caltech (DMS)

Key Issue for Your Domain

• Development of new AI algorithms and architectures for problems with ultra sparse rewards (“black swans”), e.g.
• Andrews-Curtis conjecture
• Ramsey numbers
• Diophantine equations (e.g. sum of three cubes)
• Robustness, OOD generalization
• Long horizon games

Domain-Specific Case Study/Example

• Andrews-Curtis conjecture (open since 1965)
• Slice-ribbon conjecture (open since 1962)
• Smooth Poincare Conjecture in dimension 4 (SPC4)

53

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI / machine learning for solving hard long-standing research level math problems, AI for theorem provers, AIMO challenge

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Ann Lee, Carnegie Mellon (Statistics/Machine Learning)

Key Issue for Your Domain

• Reliable scientific inference and diagnostics for settings where the likelihood is not analytically tractable but one can forward-simulate observable data.
• There is immense potential in “neuralizing” classical statistical procedures (for e.g. hypothesis testing and confidence set construction) to data settings that were previously infeasible, while achieving theoretical guarantees with a minimum of model assumptions.
• Need statisticians in close and sustained collaborations with physical scientists to realize the imminent need for user-inspired, scalable and valid statistical procedures that cross across multiple science domains

Domain-Specific Case Study/Example

• Likelihood-free frequentist inference: bridging classical statistics and machine learning for reliable simulator-based inference
• Detecting distributional differences in labeled sequence data with application to tropical cyclone satellite imagery

54

NSF AI+MPS Workshop: March 24–26, 2025; MIT

“Neuralizing” classical statistical procedures for forward and inverse problems (aka simulator-based inference) in high-energy physics, astronomy and environmental sciences

​

Founder and co-director of STAMPS@CMU

​

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Nicolás García Trillos, UW-Madison

Generative AI as a Simulation Tool in the Sciences:

• Q: How to use generative AI to conduct reliable simulations that can be used to derive trustworthy conclusions?
• Challenges:
• Pitfalls of generating far more data than originally available (i.e., theoretical obstructions).
• Off-the-shelf generative models usually overlook relevant structure.
• Does the above mean a casuistic approach to generative models in the sciences?
• Design of non-standard evaluation methods (based on specific purposes).
• Need for new protocols/guidelines for reporting results and conclusions.

Case Study: Generation of multivariate asset returns [Cetingoz and Lehalle, 2025].

• Stylized facts. E.g., marginally, heavy-tailed, asymmetric, with intertemporal dependence structure. 

• Small number of (real) data points (~7500 observations). Issue related to “model collapse” [Shumailov et al., 2024].

• Data generation for specific goals (e.g., portfolio construction): under Markowitz-like portfolio model, accuracy should not be focused on directions carrying more variance (not the standard intuition).

55

NSF AI+MPS Workshop: March 24–26, 2025; MIT

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Dima Shlyakhtenko, UCLA

Key Issue for Your Domain

• One promising direction is combining AI with proof assistants to solve math problems. However, applicability of this is highly uneven across different parts of mathematics, due to (i) varying levels of support by proof assistants and (ii) varying amounts of data available. Developing proof assistants means reproving old theorems, but may be necessary. How do we provide incentives for that?
• There are many questions, including attribution, citation, as well as “taste” with future AI-assisted math papers. Are there guardrails/best practices that are being developed across disciplines?

Domain-Specific Case Study/Example

• IPAM played a role in stimulating AI collaborations with other sciences (eg materials, electronic structure, neuroscience, etc.)
• Can learning from math (or other “hard sciences”) make better general AI? And if so, what does science get in return?

56

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Functional analysis, random matrices, free probability

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Bin Yu, UC Berkeley (Statistics, EECS, Computational BIology)

Key Issue for Your Domain

• Trustworthy uncertainty quantification in DL and genAI
• Actionable interpretations of deep learning and ML models
• Reliable and timely evaluation of genAI for safe and effective use in science and medicine
• Conceptual and theoretical understanding into deep learning models with useful insight and guidance in AI practice

Domain-Specific Case Study/Example

• Adaptive wavelet distillation from neural networks through interpretations (NeurIPS, 2021)
• LoRA+: efficient low rank adaptation of large models (ICML, 2024)
• Veridical data science: the practice of responsible data analysis and decision making (Ch. 13 on PCS-UQ) (MIT Press, 2024)
• Epistasis regulates genetic control of hypertrophy (accepted Nature Cardiovascular Research, 2024)

57

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Deep learning, PCS framework for veridical (truthful) data science, safe AI, tree-based methods, and collaborative research in neuroscience, climate science, precision medicine, and genomics.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Boris Hanin, Princeton University (DMS)

Key Issue for Your Domain

• We need new tools to describe and analyze learning with neural networks in the regime where model size, dataset size, and compute budget diverge.
• Modern ML will provide significant impetus for both conceptually interesting and practically important questions across theoretical mathematics. But to truly understand what these problems are requires a deeper integration between academia and industry.

Domain-Specific Case Study/Example

• Identifying scaling limits of neural networks – i.e. regimes where network training converges to stable but non-linear learning has proved immensely useful for practice. I have been working on doing this for modern LLMs e.g. here and in ongoing work with Cerebras (which came out of the linked paper)
• The outcome of the work with Cerebras, which we will publish soon, is significantly better performance and more efficient training, even when compared again state-of-the-art baselines, with modern LLMs.

58

NSF AI+MPS Workshop: March 24–26, 2025; MIT

​

I work on deep learning theory and seek to understand how neural networks learn and how to make them more efficient. I am also working part time at a AI compute startup called Foundry, runs a marketplace for GPUs.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Zhuoran Yang, Yale (DMS Statistics)

Key Issue for Your Domain

• Developing reliable methodologies (and/or datasets) for generative AI for better understanding which techniques work reliably and why.
• Creating more powerful tools to analyze the inner workings of AI models, with formal methods to decompose and characterize learned representations and computation patterns at a larger scale.
• Establishing principled frameworks for measuring, preventing, and addressing issues like jailbreaking, hallucinations, and unfaithful reasoning, with formal verification methods.

Domain-Specific Case Study/Example

• A theoretical explanation of how the induction head mechanism emerges from gradient-based training.
• A statistical theory of Chain-of-Thought prompting and its variants.

59

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I work on theoretical foundations of AI, particularly on reinforcement learning and language models. I aim to understand inner workings of trained transformers and how to build better LLM agents using RL.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Melanie Weber, Harvard University (DMS)

Key Issue for Your Domain

• Use of classical tools from geometry, algebra and topology to design machine learning models.
• Improving access to computing/ data infrastructure for research mathematicians to foster wider adoption of AI tools.

​

Domain-Specific Case Study/Example

• Understand the impact of data and model geometry on the learning complexity of deep neural networks.
• Design effective structured models that leverage data geometry.

60

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I work on Geometric Machine Learning, i.e., using geometric structure in data and models to design more efficient machine learning methods with provable guarantees.

​

I am also interested in AI for Mathematics.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Gianluca Guadagni, University of Virginia (DMS)

Key Issue for Your Domain

• Understand ML and AI as physical/biological systems to create new models
• Probability/Randomness is a key ingredient of ML/AI
• What kind of Math is needed in ML/AI?
• How do we teach it?

​

Domain-Specific Case Study/Example

• Train NN/AI models as random systems (statistical mechanics)

​

• Understand diffusion models and exploit conditional generative models

​

61

NSF AI+MPS Workshop: March 24–26, 2025; MIT

NN training without (or modified) Backpropagation

​

AI for Math

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI + DMS Questions to Consider

62

• What are flagship examples of AI+DMS that will help demonstrate where the field is now?
• Are the following subdomains the best way to categorize Mathematics Research?
• Mathematical Foundations
• Statistics and Combinatorics
• Applied Mathematics
• Probability
• What are the most important priorities and opportunities in AI+DMS in the next 5 years?
• What are the key challenges to address in AI+DMS?
• What does DMS have in common with the other MPS disciplines when it comes to AI?

Driving Question: How can the MPS domains best�capitalize on, and contribute to, the future of AI?

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Physics (PHY)

Breakout Discussion

NSF AI+MPS Workshop 2025

63

Feel free to add comments directly in the shared doc:

https://bit.ly/ai-mps-breakout-notes

​

Slides for report: https://bit.ly/ai-mps-breakout-report

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Your Name, Affiliation (MPS Domain)

Key Issue for Your Domain

• Describe here one key issue for your domain that you want to make sure is discussed and/or included in the white paper.
• What is the primary takeaway? What questions do you have for the group?

​

Domain-Specific Case Study/Example

• Describe here one case study or example from your experience of a success story from your domain related to AI.
• This could be a key result in research, strategies for incorporating AI, building interdisciplinary collaborations, etc.

64

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Brief description of how you’ve used AI in your work.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Physics Breakout Session Outline

1. 1:30ea introduce self/slide
2. 30s follow up comments
3. Break into 4 groups: Small group discussion
1. 20 min per topic
2. Discuss potential slide content, review anything already written
3. Add new points to slide
4. Add comments with questions about existing content on slide
4. Coffee break
5. Group discussion
6. Finalize slides together

65

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Jesse Thaler, MIT (PHY)

Key Issue for Your Domain

• Most HEP experimental research happens in large collaborations, but there are significant AI/ML innovations coming from small-author-list efforts. Besides the inherent tension between these two modes of research, there is a growing list of cases where AI/ML techniques that worked as a proof of concept face new challenges when deployed on real experimental data (e.g. anomaly detection).
• Are new modes of collaboration needed to facilitate scaling up and experimental deployment of novel AI/ML strategies? Do we need to build topical collaborations in AI/ML (analogous to lattice QCD)?

Domain-Specific Case Study/Example

• Interdisciplinary AI/ML dialogue lead to exciting connections between optimal transport (OT) and collider physics. Through conversations with MIT colleagues and a collaborations with Harvard/Tufts researchers, we "rediscovered" a half-century of particle physics techniques in the language of OT and opened a new toolbox for data analysis, interpretation, and first-principles calculations.

66

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Using techniques from optimal transport, topic modeling, simulation based inference, and point cloud learning to interpret data from the Large Hadron Collider, with a focus on QCD and jets

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Gary Shiu, U. Wisconsin-Madison (PHY)

Key Issue for Your Domain

• The availability and quality of training data in string theory limits the ability to construct AI models that can navigate the high-dimensional space for optimization of physical properties and analyzing the statistics of the string landscape. More relevant and higher quality data only partially address this problem in a huge state space of unknown size; more data-efficient methods that incorporate or discover physics principles (e.g. duality and other symmetries) are needed.
• Is there a way to encourage a concerted effort to curate and maintain a publicly accessible database that facilitates progress using AI as well as collaborations?

​

Domain-Specific Case Study/Example

• Applications of AI to string theory has mostly been carried out in scale-down versions due to limited computing resources in academia. After a chance discussion with Meta researchers after my talk at the Hammers and Nails Conference, I have initiated an ongoing collaboration in developing a transformer model for Calabi-Yau spaces.

​

• ​

67

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Using topological data analysis,simulation-based inference, and diffusion model to study cosmological data.

Using genetic algorithm, RL to find optimal string theory solutions; using transformers to generate new Calabi-Yau spaces.

​

​

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Michelle Kuchera, Davidson College (PHY)

Key Issue for Your Domain

• How can we overcome data sharing barriers/challenges across experiments or groups to make best use of foundation models / multimodal/ interoperable models / etc.
• Are these challenges stronger in some domains than others?
• How can we come together across sub-fields effectively?

Domain-Specific Case Study/Example

• We see pretrained models from one experiment using a TPC detector to provide a useful latent space for other experiments… requiring far less data for tuning for a downstream task on a different experiment. Practical issues surrounding data sharing limit the ability to scale and test this avenue further.

​

68

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Bayesian approaches to Deep Learning, stochastic modeling, point cloud and sparse tensor architectures. Diffusion models. My work is done within the context of addressing challenges in nuclear/particle physics data analysis.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Stephen Whitelam, Berkeley Lab (PHY)

Key Issue for Your Domain

• Development of optimization algorithms (for instance, related to reinforcement learning methods) for strongly fluctuating stochastic systems driven far from thermal equilibrium
• Current efforts include a variety of approaches whose benefits and limitations are not easy to compare. Can we (and should we) develop a standard set of computational systems against which to test new methods, motivated by the way ML relies on standard data sets?

Domain-Specific Case Study/Example

• Various RL methods used in the 2010s to play video games have been adapted (by us and others) to control nanoscale physical systems and experiments, identifying control strategies more effective than those derived from intuition or prior experience.

69

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Using simulation and machine learning we have developed neural-network protocols to control nanoscale process in the laboratory, including mechanical cantilevers that do logic operations and an oven than produces graphene.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Vuk Mandic, University of Minnesota (PHY)

Key Issue for Your Domain

• Robust uncertainty quantification in AI inference comes up in multiple ways in the gravitational-wave (GW) field: fast, real-time inference on GW signals is shared in alerts for EM follow-ups; robust inference on many overlapping signals and their subtraction from strain data will be necessary for future GW detectors; some GW signals come with complex statistical distributions and standard likelihood-based inference methods are not feasible - SBI approaches may be appropriate.
• Are there similar concerns across other domain fields? If so, are there approaches that could be developed/applied across fields? Would it make sense to form a broad collaboration, including CS/AI experts, around this issue?

Domain-Specific Case Study/Example

• We developed a NN to minimize the variance in the GW strain data, using a series of sensors monitoring the environment (accelerometers, microphones…) as witness channels to capture linear and non-linear couplings of noise into the strain channel. Successfully reduced noise in some parts of the frequency band.

70

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Removal of environmental contamination from LIGO time-series gravitational wave strain data.

AI methods for detecting GW signals in LIGO data, inference and uncertainty quantification.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Pankaj Mehta, Boston University (PHY)

Key Issue for Your Domain

• There is currently no single place for people studying the “Science of AI” from a physics perspective. This community tends to be scattered across statistical physics, physics of living systems, and computational neuroscience.
• How can we find a way to bring this disparate group together - through a dedicated funding mechanism, conferences, educational workshops etc.

Domain-Specific Case Study/Example

• There has been extensive literature that uses statistical physics to understanding why machine learning works despite having many more parameters than data points – both for high-dimensional regression, but also neural networks (NTK), and manifold learning.
• In a more applied setting, there is growing subfield now dedicated to using RL to understand quantum control after our initial paper in PRX in 2018.

71

NSF AI+MPS Workshop: March 24–26, 2025; MIT

statistical physics for theory of AI; pioneered use of reinforcement learning for quantum control; in physics of living systems, developed numerous techniques for understanding high-dimensional biological data

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

David Shih, Rutgers University (PHY)

Key Issue for Your Domain

• In particle physics, there has been an explosion of new proofs-of-concept for powerful new data analysis methods using modern ML. However, the new generation of researchers at the forefront of this revolution is falling through the cracks as they are not viewed as traditional theorists or experimentalists.
• How can we better break down the traditional barriers between theory and experiment in HEP? Would more open data help? How do we develop more career pathways for data-science-driven particle physicists?

Domain-Specific Case Study/Example

• We have invented a number of proofs-of-concept for new anomaly detection methods (autoencoders, ANODE, CATHODE…) that have been carried to fruition by ATLAS and CMS. A case study in lowering barriers between theorists and experimentalists yielding new results: with CMS, I worked to implement CATHODE on Run II data and this was recently made public as part of the first ever ML-powered model-agnostic search for new physics at CMS.

72

NSF AI+MPS Workshop: March 24–26, 2025; MIT

ML for fundamental physics (colliders, astro, cosmo). Anomaly detection; surrogate modeling; superresolution and upsampling; simulation-based inference; explainable AI; foundation models; fully data-driven measurements

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Murray Holland, JILA, CU Boulder (PHY)

Key Issue for Your Domain

• Reinforcement learning has proven very effective for the quantum design of unitary gates for matterwave physics. A major aspect of this work is that quantum has unique features that need to augment standard approaches; i.e., irrelevant global phase, quantum measurements with fundamental uncertainty and feedback, exponential scaling of system size requiring systematic approximations.
• Addressing these require new approaches. Are there common motivations in other areas of AI+MPS?

Domain-Specific Case Study/Example

• Using AI methods, we have developed a new approach to splitting, mirroring, and recombining the wavefunction of rubidium atoms that are Bose-Einstein condensed and manipulated in an optical lattice (crystal of light) at nanokelvin temperatures. Reinforcement learning is used to create the atom optic components. The team involves faculty from aerospace, electrical engineering, physics, and quantum information science.

73

NSF AI+MPS Workshop: March 24–26, 2025; MIT

I Use RL as a core component of a Bose-Einstein condensation experiment to perform atom interferometry for the precision measurement of inertial forces and of gravity

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Jennifer Ngadiuba, Fermilab (PHY)

Key Issue for Your Domain

A major challenge in PHY is ensuring that AI models are interpretable. Interpretability is not just a transparency issue—it is essential for robustness and trust in AI-driven decision-making at scale.

• AI models deployed in trigger and data acquisition systems must be reliable and provide clear insights into why a decision was made.�
• In anomaly detection, a lack of interpretability can lead to misclassification of rare physics signals or unknown new phenomena.�
• Developing tools and frameworks that enhance interpretability will improve model robustness and facilitate adoption in critical physics applications.

What strategies can be used to enforce interpretability without sacrificing AI performance?�How can we develop benchmarks that prioritize both accuracy and explainability in scientific AI?

Domain-Specific Case Study/Example

• HEP has successfully applied anomaly detection for new physics searches at both the trigger level and in offline analyses at the LHC. While deep learning models enhance sensitivity to rare signals, their black-box nature makes validation difficult.
• Efforts to incorporate physics-informed architectures showed that balancing explainability with performance leads to more robust and trustworthy AI applications.

74

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Photo of you

Real-time data processing systems (trigger and data acquisition), data compression, representation learning, anomaly detection.

NSF AI+MPS Workshop: March 24–26, 2025; MIT

E. Paulo Alves, UCLA (PHY)

Key Issue for Your Domain

• The challenge of accurate multi-scale modeling of plasma dynamics is at the core of some of the most important problems in plasma physics, from controlled nuclear fusion to the acceleration of cosmic rays. Reduced descriptions of plasmas that balance between accuracy and computational complexity are needed. This is of course a common theme to many other areas of science/engineering.
• Can ML be used to help discover such reduced models from the data of state-of-the-art first-principles simulations (even if these are still limited to modest spatial/temporal scales)?

Domain-Specific Case Study/Example

• Development of surrogate models of plasma turbulent transport to predict macroscopic profiles of magnetically confined plasmas [Rodriguez–Fernandez 24]. Development of moment closures for fluid models of nonlinear kinetic plasma dynamics.

75

Exploring the intersection of ML and ab initio simulations of plasmas. Used ML to uncover reduced models of plasmas from ab initio simulations; using PINN-like methods to solve inverse problems in plasma physics

NSF AI+MPS Workshop: March 24–26, 2025; MIT

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Cris Fanelli, William & Mary (PHY)

Key-issue for Your Domain

• A major barrier to AI adoption in experimental design and optimization is sociological—large-scale experimental collaborations often resist integrating AI-assisted optimization, even where it offers clear advantages. E.g., detector design and optimization remains highly compartmentalized, with individual sub-systems optimized in isolation and integrated post hoc under fixed constraints. This sequential, siloed approach fails to tackle the inherently multi-objective and high-dimensional nature of the design space, leading to suboptimal solutions.

Domain-Specific Case Study/Example

• The W&M team is developing AID²E, a scalable, distributed framework enabling AI-assisted optimization of complex detector systems. Designed to handle large design parameter spaces and multiple competing objectives with physical constraints, it facilitates holistic, global optimization strategies beyond manual or sequential methods. The framework is being applied in collaboration with the ePIC and Detector 2 working groups at the EIC, and is being deployed for several other applications beyond EIC design.
• The W&M group has also developed a RAG-based agent for EIC science, as a first step towards building agentic workflows that allow handling some complex task. A multimodal AI agent in the future could assist by interpreting visual data, cross-referencing documentation (logbooks, wiki, etc.), and providing real-time actionable insights with human-in-the-loop.

76

NSF AI+MPS Workshop: March 24–26, 2025; MIT

ML/DL-based UQ, reconstruction, high-fidelity fast sim for NP/HEP. Agentic workflows and AI-assisted design and optimization.

Photo of you

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Shih-Chieh Hsu, A3D3 / U Washington (PHY)

77

NSF AI+MPS Workshop: March 24–26, 2025; MIT

Photo of you

Director of NSF HDR Institute A3D3�Professor of Physics & Adjunct Professor of ECE, University of Washington

�Real-time AI, Exp. Particle Physics, Hardware - algorithm co-development

Key Issue for This Theme

• High Energy Physics: Focuses on understanding the fundamental nature of matter and energy through experiments like the Large Hadron Collider (ATLAS, CMS). It involves using AI for real-time data analysis, anomaly detection, and particle reconstruction to uncover new physics phenomena.
• Multi-Messenger Astrophysics: Involves studying cosmic events using multiple types of signals, such as electromagnetic radiation (ZTF), gravitational waves (LIGO), neutrinos, and cosmic rays (DUNE). AI is used for rapid detection, data integration, and analysis to provide a comprehensive understanding of astrophysical phenomena.

Domain-Specific Case Study/Example

• Low-latency AI deployment on FPGA (hls4ml)�ISFPGA 2025 proceeding MLST 4 025004 (2023)
• Inference-as-a-Service for science: (SuperSONIC)�ML4GW/HERMES, Nature Astronomy 6, 529–536 (2022) Phys. Rev. D 111, 042010 (2025) arxiv:2501.05520
• Neurosciences: real-time LFADS on FPGA arxiv:2402.04274, ICCAD 2023

PHY-2117997

NSF AI+MPS Workshop: March 24–26, 2025; MIT

AI + PHY Questions to Consider

78

• What are flagship examples of AI+PHY that will help demonstrate where the field is now?
• The core research areas in NSF PHY are as follows. Is this the right organization for the AI white paper? How to ensure coverage across all of these areas?
• Atomic, Molecular and Optical Physics
• Elementary Particle Physics
• Gravitational Physics
• Nuclear Physics
• What are the most important priorities and opportunities in AI+PHY in the next 5 years?
• What are the key challenges to address in AI+PHY?
• Do these answers differ between large-scale experiments, small-scale experiments, and theory?
• What does PHY have in common with the other MPS disciplines when it comes to AI?

Driving Question: How can the MPS domains best�capitalize on, and contribute to, the future of AI?

• Particle Astrophysics and Cosmology
• Physics of Living Systems
• Plasma Physics
• Quantum Information Science

NSF AI+MPS Workshop: March 24–26, 2025; MIT