Abstract: TDeep neural operators have emerged as a powerful paradigm in scientific machine learning (SciML) for learning nonlinear mappings between function spaces arising in complex dynamical systems. In this talk, I will introduce the deep operator network (DeepONet) and its extensions, such as geometry-dependent/manifold operator learning, and demonstrate their effectiveness on diverse multiphysics and multiscale 3D problems, such as geological carbon sequestration, fire simulation, full waveform inversion, and topology optimization. I will then present the first operator learning method requiring only a single PDE solution, i.e., one-shot learning, enabled by a new concept of local solution operator motivated by the locality of physics. Moreover, I will discuss our recent work on diffusion models, including FunDiff as a novel framework of diffusion models over function spaces for physics-informed generative modeling, and RED-DiffEq which leverages denoising diffusion models as regularization for inverse PDE problems. Finally, I will present a physics-informed multimodal foundation model for PDEs and neural-operator element method (an efficient and scalable finite element method enabled by reusable neural operators).
Abstract: The goal of the workshop is to bring together experts in geometric and combinatorial group theory, low-dimensional topology, and orderable group theory to communicate recent results concerning the normal rank (weight) of groups, Dehn surgery numbers of 3-manifolds, and related problems. The normal rank of a group is the smallest number of elements that normally generate the group. This is a mysterious notion, and several longstanding problems center around it (such as the Kervaire conjecture, Levin conjecture, and the Gordon conjecture). The Dehn surgery number for 3-manifolds is bounded below by the normal rank of the fundamental group. We hope that the workshop will provide a stimulating environment for participants to exchange ideas, discuss recent advances, and draw inspiration from one another in light of the latest progress on these topics.
Speaker: Bill Regli (UMD) - https://me.umd.edu/clark/faculty/902/William-Regli When: Tue, December 1, 2026 - 2:00pm Where: Kirwan Hall 3206
View Abstract
Abstract: This talk explores some of the practical issues surrounding the implementation of neurosymbolic systems with a specific focus on if and how they could be realized using alternative computing substrates.
Computing as a general-purpose fabric for mapping among different forms of information has progressed beyond the transistors and VLSI circuits etched in a traditional silicon CPU or GPU. Modern AI models are trained and executed on (essentially) massive parallelized linear algebra supercomputers. But future neurosymbolic systems have an inherently hybrid architecture potentially ammeanble to new, alternative, computing substrates.
Speaker: Sarah Wiegreffe (UMD) - https://sarahwie.github.io/ When: Tue, December 1, 2026 - 2:30pm Where: Kirwan Hall 3206
View Abstract
Abstract: Large language models (LLMs) power a rapidly-growing and increasingly impactful suite of AI technologies. However, their scale and complexity leave us without a fundamental scientific understanding of much of their behavior, or of how to modify it on the fly. In this talk, I will describe some of our recent work on (mechanistically) interpreting and steering LMs. This includes 1) interpreting latent reasoning models and 2) understanding the mechanisms driving inference-time steering methods for refusal.
Speaker: Nicholas Nelsen (UT Austin) -Scientific data increasingly takes the form of evolving point cloud ensembles, interacting particles, or varying-size collections of sensor measurements, rather than fixed-dimensional data representations. These data arise in applications such as probabilistic weather forecasting, LiDAR imaging, fusion energy science, and multi-agent control. This talk outlines a framework for machine learning on the space of probability distributions, which treats entire datasets as the input or output object of interest. It discusses two complementary formulations: neural operators acting on probability density functions and transformers acting on probability measures through particle representations. Toward analyzing the capabilities and limitations of such architectures, the talk presents recent approximation, stability, and discretization theory that underpin their computational cost and accuracy. A running motivation throughout the talk is amortized Bayesian inference, where the goal is to approximate probabilistic conditioning across a family of joint distributions in near real-time using a single trained model. When: Tue, December 1, 2026 - 3:30pm Where: Kirwan Hall 3206