Enhancing surrogate models of engineering structures with graph-based and physics-informed learning

Abstract

This thesis addresses several opportunities in the development of surrogate models used for structural design. Though surrogate models have become an indispensable tool in the design and analysis of structural systems, their scope is often limited by the parametric design spaces on which they were built. In response, this work leverages recent advancements in geometric deep learning to propose a graph-based surrogate model (GSM). The GSM learns directly on the geometry of a structure and thus can learn on designs from multiple sources without the typical restrictions of a parametric design space.

Engineering surrogate models are often limited by data availability, since designs and performance data can be expensive to produce. This work shows that transfer learning, through which training data of varying topology, complexity, loads and applications are repurposed for new predictive tasks, can be used to improve the data efficiency of surrogates, often reducing the required amount of training data by one or two orders of magnitude. This work also explores new potential sources for training data, namely engineering design competitions, and presents SimJEB, a new public dataset of simulated engineering components designed specifically for benchmarking surrogate models. Finally, this work explores the emerging technology of physics-informed neural networks (PINNs) for structural surrogate modeling, proposing two new heuristics for improving the convergence and accuracy of PINNs in practice. Combined, these contributions advance the generalizability and data efficiency of surrogate models used in structural design.