GeoPIPN: A Novel Geometry-Aware Deep Learning PDE Solver for Faster Thermal Quality Control on Unseen Geometries

Abstract

This presentation proposes Geometry and Physics Informed PointNet (GeoPIPN), a novel deep learning-based Partial Differential Equation (PDE) solver tailored for quality control and reliability engineering by rapidly solving PDEs across unseen geometries without further training. PDEs model physical processes like thermal distribution, crucial in additive manufacturing, where precise thermal control ensures desired material properties, uniform microstructure, and dimensional accuracy – key to product integrity and durability. Traditional numerical PDE solvers are often too slow to support the fast-paced, iterative design phase, while classical Physics Informed Neural Networks (PINNs) require retraining for each new geometry, making them inefficient for continuously evolving designs. GeoPIPN overcomes these limitations of generalizing model across different geometric configurations by leveraging a combination of geometry-aware global and boundary features along with local features. Through two case studies, each involving distinct geometric domains governed by thermal distribution physics, we demonstrate significant superior performance of GeoPIPN compared to current state-of-the-art benchmark. GeoPIPN thus offers a powerful solution for quality control and reliability engineering, enabling consistent, repeatable simulations across varying geometries and advancing PDE-driven design to support high-quality, reliable manufacturing processes.