PhD Final Defense – Weiheng Zhong

Physics-AI Algorithms Across Diverse Data Regimes:  Sensors, Simulations, and Videos

Advisor: Professor Hadi Meidani

Abstract

Artificial Intelligence (AI) is rapidly transitioning into the physical world, driving the emergence of Physics AI—a domain dedicated to training intelligent agents that predict the evolution of physical systems while strictly respecting physical laws. Because data forms the foundation of modern architectures, this dissertation addresses the escalating technical and cognitive hurdles of Physics AI across three core training modalities: sensors, simulations, and videos. To overcome these challenges, this work introduces novel, physics-respecting algorithms, demonstrating that models embedded with deep physical understanding consistently yield distinct advantages over those with less.

Within the sensor data regime, we target both dense and sparse configurations to demonstrate how embedding physical knowledge compensates for varying data availability. For data-dense environments, we introduce architectures that simultaneously learn empirical patterns and physical laws, yielding superior predictive performance alongside critical, trust-enhancing interpretability. Conversely, for sparse configurations requiring real-time field reconstruction, we employ a physics-informed variational autoencoder framework. This approach matches the rapid execution efficiency of standard interpolation while delivering robust uncertainty quantification, underscoring how physical priors can successfully reconstruct full system dynamics from highly insufficient data.

Expanding the scope of Physics AI beyond physical instrumentation, the simulation data regime leverages numerical environments to eliminate reliance on hardware sensors. For idealized problems with simplified boundary assumptions, we implement completely physics-informed, data-free training methodologies to predict partial differential equation (PDE) solutions directly, demonstrating that physical laws can effectively substitute for expensive synthetic datasets to minimize computational overhead. Transitioning to complex, detailed design evaluations, we develop specialized architectures across distinct geometric representations, proving that tailored input features substantially enhance an AI model's capacity to respect underlying physical constraints and scale efficiently. Finally, to resolve Out-of-Distribution (OOD) generalization challenges, we utilize a hybrid algorithm that couples synthetic data pre-training with sample-based physics fine-tuning to ensure robust deployment safety and reliability.

Finally, within the video data regime—where unconstrained environments are computationally intractable to replicate via traditional simulation—we leverage video streams to train models capable of capturing generalized physical patterns. We design a novel framework by coupling a generative video foundation model with a high-performance, GPU-accelerated physics solver; this integration demonstrates the potential of enforcing physical constraints within modern generative architectures to substantially increase the physical plausibility and trustworthiness of future-state predictions.