Towards End-to-End Prompt-Vision-Physics Neural Network for Fast Design Discovery
Abstract
In this paper, we present the study of an end-to-end prompt-vision-physics neural network (PVPNN) to speed up the design discovery process in science and engineering. PVPNN has the potential to transform design discovery by enabling researchers and engineers to specify design requirements seamlessly via prompts, analyze available large-scale relevant historical data sets to generate novel physics-conform designs and predict the efficacy of potential design candidates rapidly. Starting from a given image or prompt for an ideal design as the baseline, our proposed framework first generates potential designs from a vision foundation model. Based on the practicality requirements provided by the user through text prompt, the PVPNN initializes a resource-efficient physics-informed neural network and jointly optimizes both geometry and physics models in an end-to-end mode. An experimental study on a simple engineering structure validates that our proposed framework can seamlessly satisfy visual preferences and practicality for fast design discovery.
Authors: Qingshan Xu, Jiao Liu, Melvin Wong, Ge Jin, Ryan Lau, Yew-Soon Ong, Stefan Menzel, Thiago Rios, Joo‐Hwee Lim, Chin Chun Ooi
Published in: IEEE Conference on Artificial Intelligence (CAI) (2024)