Neocortex Seminar Series
Physics-Aware AI at Scale: Neural Compression and Vision Transformers for Simulation Data
April 15, 2026
2:00 pm – 3:00 pm Eastern time
Speaker: Jessica Ezemba, Carnegie Mellon University
Abstract: Modern scientific computing generates terabyte-scale simulation data across physics domains, yet researchers lack efficient tools for storage, retrieval, and analysis. This work presents two complementary approaches to addressing this bottleneck, both leveraging wafer-scale computing on the Cerebras CS-3. First, SINCPS (Semantic-aware Implicit Neural Compression for Physics Simulations) compresses physics simulation data by 150× to 25,000× using implicit neural representations, reducing per-dataset training time to 2 to 3 hours while preserving physics-critical conservation laws across 22 datasets from The Well benchmark. Second, PhySiViT, a domain-specific Vision Transformer trained on approximately 7 million physics simulation images from The Well in just 22 hours, produces embeddings with distinct physics-informed structure that outperforms general-purpose models like CLIP and DINOv2 on physics-specific tasks, achieving 43% better temporal forecasting (R² = 0.33 vs. 0.23) and superior physics clustering (silhouette score = 0.23 vs. 0.20). Together, these systems form a pipeline for transforming large simulation archives into compact, queryable representations suitable for downstream machine learning workflows, demonstrating that domain-focused models trained efficiently on specialized hardware can outperform general-purpose counterparts on scientific tasks.
Speaker Bio: Jessica Ezemba is a Ph.D. candidate in Mechanical Engineering at Carnegie Mellon University, where she is advised by Dr. Christopher McComb and Dr. Conrad Tucker. Her research sits at the intersection of artificial intelligence and engineering design, with a focus on making the design process faster and less prone to errors. She is particularly interested in engineering simulations, a critical yet time-consuming and expertise-dependent stage of design where products are tested before manufacturing. Her work investigates how AI can accelerate simulation interpretation by enabling faster integration of multidisciplinary expertise. Through developing benchmarks, foundation models, and surrogate modeling approaches, Jessica has demonstrated that current AI tools struggle to understand engineering simulations, motivating her ongoing work on alternative paradigms, including agentic workflows that keep humans in the design loop while enabling automated understanding. She has leveraged wafer-scale computing on the Cerebras CS-3 to develop physics-focused foundation models and neural compression methods for large-scale simulation data, including PhySiViT, a domain-specific Vision Transformer for physics simulations, and SINCPS, an implicit neural compression framework achieving up to 25,000× compression on terabyte-scale simulation archives. She has collaborated with industry partners including Ansys, and her research has been published in venues including the ASME Journal of Mechanical Design, and the ACM/IEEE Supercomputing Conference (SC).