Session 1:

1:00 to 2:00 p.m.

“AI-Enable Surrogate Models for Accelerating Scientific Computing” 

David Radice, “Accelerating and Improving Neutron Star Merger Simulations with Emulators”:

• Abstract: Neutron star mergers are among the most violent events in the Universe. The interpretation of the rich gravitational-wave and electromagnetic radiation they produce requires computationally intensive, multi-physics, multi-scale simulations. This talk presents ongoing work in my group to accelerate aspects of these simulations and improve their fidelity using scientific machine learning. In particular, I will discuss a new approach to accelerate neutrino microphysics calculations and embed unresolvable, small-scale physics in global simulations using emulators.

Steven Greybush, “Building, Applying, and Evaluating Surrogate Models for Weather Prediction and Data Assimilation”:

• Abstract: In this talk, I will discuss how AI surrogate models are used for weather prediction and data assimilation.  The talk will survey recent research, as well as discuss challenges for the field.  Examples include building a surrogate model for predicting the formation of new thunderstorms from satellite data, evaluating emulators for numerical weather prediction of winter storms, and using AI surrogates to generate large ensembles for data assimilation, which is the process of estimating the state of a dynamical system and its uncertainty.

Ashwin Renganathan, “Sample-efficient and optimal decision-making with probabilistic surrogate models for aerospace design”:

• Abstract: High-fidelity simulations are increasingly central to aerospace design and analysis, but their computational expense can make optimization, uncertainty quantification, and reliability assessment prohibitively costly. This talk explores probabilistic surrogate models as tools for sequential decision-making. For multiobjective design, we will discuss Pareto-optimal Thompson sampling (qPOTS), which combines Gaussian-process surrogates with posterior sampling to efficiently identify batches of promising Pareto-optimal designs, and recent extensions that exploit correlations between objectives and constraints to decouple expensive oracle evaluations and selectively acquire only the most informative quantities. I will then consider reliability analysis, where sample efficiency requires concentrating computational effort near rare but consequential failure regions. This includes deep Gaussian-process surrogates for nonstationary transonic aerodynamic responses and adaptive multifidelity methods that jointly determine where to sample and at what simulation fidelity, coupled with importance sampling for rare-event estimation. Across these applications, the central theme is that the greatest value of probabilistic surrogates lies in using their uncertainty to determine what computation to perform next, enabling reliable design decisions with substantially fewer high-fidelity simulations.

Session 2:

2:10 to 3:10 p.m.

“AI-Enable Surrogate Models for Accelerating Scientific Computing“

Xiaofeng Liu, “Physics-Informed Operator Learning for Computational Hydrodynamics: What the Physics Buys, What It Costs, and Remedial Options”:

• Abstract: Two-dimensional hydrodynamic solvers are the workhorses of flood simulations and river/coastal engineering. This talk presents a physics-informed Deep Operator Network that learns the solution operator of the 2D shallow water equations. Trained on a physics-based solver simulation data of a Sacramento River reach, the surrogate predicts a new scenario in under a second. The embedded PDE residual makes out-of-distribution error grow about four times more slowly, at a small cost in in-distribution accuracy. That cost is structural: the pointwise residual asks the network to satisfy the continuous equations, while the training data are a discrete finite-volume solution that does not, so the two objectives contend and neither is met exactly. I close with finite-volume-informed networks, which replace the pointwise residual with a differentiable Roe flux balance to enforce conservation at the control-volume and domain level instead.

Daning Huang, “Reduced-order modeling with agentic AI for designing dynamical systems”:

• Abstract: Complex engineered dynamical systems—from aircraft and nuclear reactors to power grids—can be high-dimensional in states, controls, and design variables. These dimensions compound in simulation-based design and optimization, often making direct high-fidelity analysis computationally prohibitive. Reduced-order models can alleviate the state dimension, but many conventional approaches do not generalize well across parameters.  In this talk, we present a reduced-order modeling framework that leverages known physical structure and the geometry of the data to achieve both dynamical accuracy and parametric generalization. We also describe an existing agentic AI workflow for ROM construction and analysis. Building on this capability, we discuss a broader vision in which agentic AI connects ROMs with design exploration and optimization, enabling increasingly scalable and autonomous design of complex dynamical systems.

Samuel Grauer, “Latent dynamics for data assimilation in flows with shocks”:

• Abstract: Flows containing quasi-discontinuous features, such as shock waves and flames, are common in power and propulsion applications. Predictive simulations are essential for design, but they are challenging due to their computational cost, the numerical difficulties of capturing evolving discontinuities, and the limitations of current physical models. Data assimilation (DA) combines measurements with numerical simulations to estimate the evolving flow state and unknown system parameters, and it can also be used to specify low-cost surrogate models for design. However, flows with shocks and flames pose distinct challenges for conventional DA methods. This talk will briefly examine these challenges and explore the use of neural surrogate models for DA. It will show how jointly training an autoencoder and a neural ordinary differential equation shapes the latent space to support effective assimilation. The resulting decoder and latent flow map are then embedded within a DA algorithm. Examples will be presented for 1D and 2D Riemann problems.