Date: November 5 to 7, 2026
Location: Westin, Arlington
Address: 801 North Glebe Road,
Arlington, Virginia 22203
Room: Hemingway Salon 1
Scientific discovery is increasingly constrained not by the lack of simulation tools, but by the immense computational, data, and energy costs required to explore complex scientific systems at scale. Modern AI methods have demonstrated extraordinary capabilities in large-data regimes, yet many frontier scientific domains—such as climate modeling, fluid dynamics, materials discovery, fusion, subsurface systems, and biological modeling—operate in precisely the opposite setting: sparse observations, expensive simulations, limited experimental access, and strict reliability requirements.
The primary goal of the symposium is to unite researchers working on efficient, physically grounded, and resource-aware AI methods for scientific discovery. The symposium will highlight emerging advances in areas such as data-efficient learning, hybrid neuro-mechanistic modeling, scientific foundation models, uncertainty-aware AI, operator learning, and compute-aware reasoning for scientific applications. Ultimately, the symposium seeks to catalyze a new research community around parsimonious scientific machine learning (Parsimonious SciML) and establish a roadmap for scalable scientific AI that prioritizes efficiency, trustworthiness, and scientific utility rather than brute-force computation alone. This bridge builds upon the success of three previous symposia organized on a related topic of Knowledge-Guided Machine Learning at the AAAI bridge program in 2024, 2025 and 2026. Accepted papers will appear in AAAI proceedings.
Day 1: November 5
10.30am - 11.00am
11.00am - 11.30am
Abstract: Machine learning has opened new possibilities for molecular and materials simulations, from accelerating expensive quantum-mechanical calculations for property prediction to designing new materials with desired properties. Unlike many applications of machine learning, however, atomistic ML is accompanied by theories that hint at the structure of the problem and the relationships between the quantities one seeks to model. In this talk, I will discuss how this knowledge can serve as a compass for navigating this rapidly evolving ML landscape and for making the most of available data. I will present examples across scales, from electronic structure and molecular mechanics, showing how physical principles can guide the design of model inputs and architectures, the selection of what we predict and constrain, and the integration of ML within simulations.
Bio: Dr. Jigyasa Nigam is an Associate Research Scientist at the Flatiron Institute’s Initiative for Computational Catalysis. Before joining the Flatiron Institute, Jigyasa was an Engineering Excellence postdoc fellow at MIT and earned her PhD in Physics from EPFL in 2024. Her research focuses on developing atomistic machine learning methods and bridging machine learning with first-principles methods to accelerate computational chemistry and materials discovery.
11.30am - 12.00pm
Abstract: In the past decade self-driving materials and chemistry laboratories have advanced from exploratory research concepts to an increasingly mature technology being deployed across many areas of science. At the current moment this field is limited by needed advances in scientific machine learning that can improve automated capabilities of obtaining compact and structured scientifically relevant information from complex multimodal data. This talk will give an overview of some of these opportunities in the context of a nascent automated synchrotron testbed for studying dynamic structural evolution of metal alloys during the highly non-equilibrium conditions induced by laser melting in metal additive manufacturing process. We present novel algorithms for extracting symbolic representations of the evolving crystal structures from the raw data. The phase evolution behavior determined by these measurements drives a dynamic experiment planning loop that uses active learning methods to efficiently map solidification kinetics information as a function of alloy composition and processing. We will discuss the potential acceleration for alloy and processing design and the opportunities for fusing this measurement capability with other structural and property characterization modalities. This research highlights important aspects of the challenges in automated characterization and symbolic hypothesis representation that are needed to advance self-driving laboratory systems from optimization harnesses to scientific hypothesis evaluation machines.
Bio: Brian DeCost is a materials scientist in the Material Data Division at the National Institute of Standards and Technology. He earned a B.S. in Chemical Engineering at the University of Florida and a Ph.D. in Materials Science and Engineering at Carnegie Mellon University. Brian’s research focuses on developing and applying scientific machine learning methods to build intelligent measurement systems that address fundamental and applied problems in microstructure science and alloy design.
12.00pm - 12.30pm
Abstract: Machine learning is playing an increasingly vital role in physics modeling. On this evolving frontier is so-called physics-aware deep learning (PADL). Unlike conventional deep neural networks that focus on an input-output mapping, PADL algorithms can make accurate predictions of physical phenomena, while adhering to physical principles and constraints. So far, on relatively simple spatiotemporal dynamics problems, such as predicting vortex shedding in Navier-Stokes simulations, the prediction accuracy and speed of PADL algorithms have been demonstrated. However, when it comes to problems with strong nonlinearity, characterized by sharp gradients, fast-transient features, discontinuous evolving boundaries, etc., common in many defense-related problems, PADL’s effectiveness still remains unclear. Particularly, from a purely statistical perspective, these extreme features are viewed as “statistical outliers” that must be disregarded to prevent overfitting; while from the physics standpoint, these are in fact the most important characteristics that must be modeled precisely. Such a gap between the statistical view of model training and the scientific needs for predicting nonlinear features renders an interesting opportunity for transdisciplinary research, which has been the main focus of my research team lately. In this talk, I will provide an overview of the mathematical and computational challenges in applying modern deep learning models for solving real-world physics problems. I will then introduce our recent work called physics-aware recurrent convolutional neural networks (PARC) and show its effectiveness in modeling extreme physical phenomena. I will also compare PARC with other established PADL baselines and discuss current challenges and future directions.
Bio: Dr. Stephen Baek is an applied geometer and data scientist with a keen interest in geometric data analysis. He applies computational (differential) geometry, computer vision, and machine learning to address problems in engineering, physics, materials science, and medicine. He earned his bachelor’s and PhD degree in Mechanical and Aerospace Engineering from Seoul National University, Korea. He was a postdoctoral researcher at the Institute for Advanced Machinery Design in Seoul National University. Dr. Baek was the recipient of the National Science and Engineering Scholarship and the Global Ph.D. Fellowship from the Korean Ministry of Education, and the Presidential Postdoc Fellowship from the President of the Republic of Korea. Currently, he holds an appointment at the University of Virginia as Professor of Data Science. He has been leading various research projects sponsored by government and industry sponsors, such as the NSF, NIH, Air Force, Army, NASA, DOT, and Hyundai Motors.
12.30pm - 2.00pm
2.00pm - 3.00pm
RP: Extrusion Networks to Overcome the Curse of (Simulation) Dimensionality.
Exploring Joint-Embedding Predictive Architectures for Self-Supervised Representation Learning of Powder X-ray Diffraction Patterns.
On challenges in scientific knowledge representation and active learning in a self-driving additive manufacturing testbed.
RP: Auditing Process Knowledge for Parsimonious and Explainable Industrial Cyberattack Detection.
RP: Sparse-Context Science using Conditional Neural Processes.
RP: Beyond One Pixel: Parameter-Efficient Spatial Context Adaptation for Hyperspectral Image Classification.
RP: AeroGNN: A Multi-Fidelity Physics-Informed Graph Neural Network for Aerodynamic Design Optimization of Tall Buildings.
A Trustworthy Multimodal Learning Framework for Predicting Lipid Nanoparticle–RNA Delivery Performance.
RP: Conditioning GMFlow with a Gated Cross-Correlation Prior for Particle Image Velocimetry.
Characterizing LLM Code Generation for Computational Simulations.
3.00pm - 3.30pm
RP: Extrusion Networks to Overcome the Curse of (Simulation) Dimensionality
Authors: Harsha Vardhan Dasari, Arya Vaidya, Shital Adhikari and Nikhil Muralidhar.
RP: Sparse-Context Science using Conditional Neural Processes.
Authors: Emily Shinkle, Prescott Alexander, Nicholas Lubbers and Yulia Pimonova.
RP: Beyond One Pixel: Parameter-Efficient Spatial Context Adaptation for Hyperspectral Image Classification.
Authors: Siqi Du, Deepak Kandel and Dimah Dera.
3.30pm - 4.00pm
4.00pm - 5.00pm
Day 2: November 6
9.00am - 9.45am
Abstract: Dynamic graph representations, with structured labels on the graph nodes, enable parsimony in scientific modeling both at the level of the system state and at the level of the mathematical model statement. I will exhibit examples from biology (botany and neuroscience) and geoscience (subsurface fluid flow). For parsimony of system state, the key activity is approximate model reduction i.e. creation of surrogate models. Hence the requirement for a symbolic modeling language that operates on labelled graph representations and is closed under common scientific model reductions. “Dynamical Graph Grammars” are well-founded mathematically using operator algebra in the Master Equation of nonequilibrium statistical mechanics, and provide natural expression of complex multiprocess dynamics in terms of quantitative rules. I will outline the mathematics of DGGs and of parsimonious “skeletal graph products”, and some algorithmic and software tools. I will discuss the use of graph dynamics for model reduction in biological modeling and subsurface flow, the latter case focussing more on machine learning than on DGG representation which is however natural for the domain.
Bio: Eric Mjolsness is an active Professor Emeritus at the University of California Irvine in the Computer Science Department with joint appointments in the Departments of Mathematics and Systems Biology, and with the Los Alamos National Laboratory. He is a Fellow of the American Association for the Advancement of Science. He received his PhD from Caltech in Physics and Computer Science in 1986, where he was advised by John Hopfield and TA’d for Richard Feynman. His experience includes working at the Jet Propulsion Laboratory, UCSD, and the Yale Department of Computer Science. Mjolsness has research interests in AI/ML for biological modeling, including multiscale scientific models in cellular, developmental, and neurobiology, and geoscience. Current efforts center on machine learning for model reduction, together with verifiable symbolic AI model transformations, in the setting of dynamical graph grammars, for large-scale computational science. He co-initiated the NSF-funded “Computable Plant” project, and recently co-initiated the “Geophysics.ai” project at UC Irvine.
9.45am - 10.30am
Abstract: Charles Darwin described the geologic record as "a history of the world imperfectly kept, and written in a changing dialect." We describe Transparent Earth 2 (TE2), a foundation model of the Earth's subsurface that speaks these dialects and translates them into actionable information. TE2 is a single global foundation model for the Earth's subsurface that performs over 200 tasks using parsimonious data in-context at inference time. Many subsurface geoscientists hold the view that region-specific and task-specific artificial intelligence (AI) models will outperform global or multi-task AI models. TE2 demonstrates that this widely held view is wrong by outperforming multiple task-specific and region-specific models across diverse applications including bathymetry, lithology, and geothermal use-cases.
Bio: Daniel O'Malley leads the Artificial Intelligence for Geoscience Team at Los Alamos National Laboratory (LANL). His research interests include artificial intelligence, quantum computing, and computational science. He earned a B.S. in computer science and mathematics, an M.S. in mathematics, and a Ph.D. in applied mathematics, all from Purdue University. He has received numerous awards including the DOE Early Career Award, LANL Early Career Research award, LANL Director's Postdoctoral Fellowship, InterPore-Fraunhofer Award for Young Researchers, Charles C. Chappelle Fellowship from Purdue University, and LANL Large Team Distinguished Performance Award. He has published more than 100 peer-reviewed articles.
10.30am - 11.00am
11.00am - 11.30am
Geotechnical Parameter Identification beyond Force Reconstruction
Author: Honglai Peng, Rodrigo Borela, Kimberlie Staheli and J. David Frost.
RP: Multi-Fidelity Surrogates for Gappy, Expensive Simulation Data in Gravitational-Wave Astronomy
Author: Tousif Islam.
EC: Provable Extrapolation for Neural PDE Surrogates.
Author: Ignacy Nieweglowski.
11.30am - 12.30pm
Charging the Prompt to the Budget: LLM-Proposed Descriptors Reduce Selection Risk, Not Error.
RP:SAGE: Self-Evolving Acquisition through Guided Escalation for Subtle Anomaly Segmentation.
Geotechnical Parameter Identification beyond Force Reconstruction.
RP: DLens: Closed-Loop AI Agents for Parsimonious Scientific Machine Learning in Dark Matter Gravitational Lensing.
RP: Multi-Fidelity Surrogates for Gappy, Expensive Simulation Data in Gravitational-Wave Astronomy.
A Constrained Semi-Supervised Learning Algorithm for Deforestation Mapping.
Classifier-pruned Bayesian optimization for particle accelerator tuning: Exploring temporally structured manifold of 6D beam phase space.
RP: IDENTIFY: Conditional Finite-Sample Law-Set Certificates and Ambiguity-Directed Experiments.
RP:Neural ODEs with Invariants by Construction.
DF: Fidelity and Efficiency in Control-Theoretic On-the-Fly Learning for Molecular Dynamics.
12.30pm - 2.00pm
2.00pm - 2.30pm
DF: Fidelity and Efficiency in Control-Theoretic On-the-Fly Learning for Molecular Dynamics.
Author: Michael Tynes.
Charging the Prompt to the Budget: LLM-Proposed Descriptors Reduce Selection Risk, Not Error.
Author: Varshitha Manjunath and Srimonti Dutta.
RP: DLens: Closed-Loop AI Agents for Parsimonious Scientific Machine Learning in Dark Matter Gravitational Lensing.
Authors: Aatmaj Amol Salunke, Mywish Anand, Pranath Reddy Kumbam, Michael W. Toomey and Sergei V. Gleyzer.
2.30pm - 3.00pm
Abstract: The exponential growth of Earth-observation archives makes data-efficient training essential. Yet standard data pruning methods were designed for balanced classification, and they break down in the imbalanced, geographically structured settings that define geospatial AI. This talk presents three advances that address this gap. First, we show that existing dynamic pruning methods such as InfoBatch fail silently on imbalanced remote sensing segmentation. Overall accuracy remains stable, but minority-class detection collapses; for example, dead trees account for only 0.37% of pixels. We introduce Class-Stratified InfoBatch (CS-IB), which restores minority-class F1 to above-baseline levels at a 90% prune ratio while cutting training time by 38%. Second, we conduct the first systematic study of data pruning for multi-temporal change detection. We find that no existing criterion reliably outperforms random selection, because change detection requires preserving change-distribution fidelity. That is, a coreset’s changed-pixel ratio must match that of the full dataset, a property current methods lack. Our Fidelity-Diversity-Consistency (FDC) framework operationalizes this insight and achieves consistent gains, especially at extreme 1% retention budgets. Third, we demonstrate that pool-wide selection distorts the geographic and foreground-prevalence structure of remote sensing datasets under severe pruning. We propose STAC (STrAtified Coverage), which allocates the selection budget across data-defined strata. Within each stratum, STAC optimizes multi-view facility-location coverage by combining image appearance, annotation morphology, and early model behavior. It achieves state-of-the-art mIoU at retention rates of both 10% and 1% for forest segmentation. It also reveals that fine-tuning a pretrained foundation model on just 1% of labeled data can match full-dataset performance.
Bio: Raju is a Chancellor’s Faculty Excellence Program Cluster Professor of Geospatial Analytics in the Department of Computer Science at North Carolina State University (NCSU). His research focuses on Geospatial AI, with applications in national security, geospatial intelligence, natural resources, climate change impacts, and real-time actionable knowledge extraction. Before joining NCSU in 2014, he was the first lead data scientist at Oak Ridge National Laboratory, where he played a leadership role in establishing a data science research thrust within the Computational Sciences and Engineering Division. He has published more than 100 peer-reviewed articles in leading conferences (e.g., KDD, ICDM, SDM, ECML/PKDD, ACM SIGSPATIAL GIS, NeurIPS) and journals (e.g., TPAMI), and has edited two books on Knowledge Discovery from Sensor Data. He has served as an Area Chair and as a Senior Program Committee (SPC) member and Program Committee (PC) member for major international conferences (IJCNN, KDD, ICDM, WSDM, ECML/PKDD, PAKDD, CIKM, AAAI, IJCAI). He is a PC Co-chair for IEEE Big Data 2026, General Co-chair for SSTD 2027, and PC Co-chair for BDA 2027. He holds an MS and a Ph.D. in computer science from the University of Minnesota.
3.00pm - 3.30pm
Abstract: Training modern AI models increasingly requires optimizing not only model parameters but also the learning process itself, such as which data to train on and which hyperparameters to use. Such hierarchical (bilevel) learning can yield much better models under the same compute budget, but the training process is fragile and scaling it to foundation models is challenging. In this talk, I will present our work addressing this challenge from three complementary perspectives. First, I will discuss bilevel optimization theory beyond standard smoothness assumptions, with nearly optimal complexity guarantees. Second, I will introduce BLISS, a lightweight bilevel data selection framework that uses small proxy models to reduce the cost of optimizing training data for foundation models. Third, I will present FlashBoB, an I/O-efficient GPU algorithm that enables exact higher-order differentiation through attention without materializing quadratic intermediate tensors. Together, these works point toward scalable hierarchical learning by bridging mathematical foundations, efficient learning algorithms, and hardware-aware computation.
Bio: Mingrui Liu is an Assistant Professor of Computer Science at George Mason University. He received his Ph.D. in Computer Science from the University of Iowa in 2020 and was a postdoctoral fellow at Boston University.
His research lies at the intersection of machine learning, mathematical optimization, statistical learning theory, and efficient AI systems. He has published over 30 papers in leading venues, including NeurIPS, ICML, ICLR, COLT, and JMLR, and has served as an Area Chair for major machine learning conferences. His research is supported in part by NSF, including a Mathematical Foundations of Artificial Intelligence (MFAI) award. He was selected for the AAAI 2024 New Faculty Highlights program and received the 2024 Cisco Faculty Research Award.
3.30pm - 4.00pm
4.00pm - 5.00pm
Day 3: November 7
10.30am - 11.00am
11.00 - 11.45am
11.45 - 12.30pm
Title: Parsimonious Fine-Tuning of Scientific Foundation Models that Predict Time Series Evolution of Partial Differential Equations
12.30 - 12.40pm
Organizing Committee
Steering Committee