Physics-Informed Machine Learning for Many-Objective Generative Design

Overview

Generative design (GD) is a computational method for the exploration of optimized engineering structures based on machine learning. Although well suited for exploring designs with various objectives, GD is often only used to generate parts which are stiff yet lightweight, ignoring the unique goals of design problems. The use of physics equations and intuition in the design of machine learning models, or physics-informed machine learning (PIML), has been demonstrated to improve machine learning models. Explo iting GD’s potential for many objective design exploration, the present study proposes PIML as a method of creating a many -objective GD (MOGD) program. This proposition is demonstrated through a case study in an automotive wheel design problem, generating strong, thermally dissipative, lightweight, and diverse designs. A physics-informed neural network is developed to generate wheels based on input parameters, which are selected based on physics principles to be impactful to the performance of designs. Then, a many -objective reinforcement learning model is trained to generate sets of input parameters for the generator with sensitivity to an engineer’s input regarding the priority to be placed on each objective. Connecting these models yields an MOGD program which can generate a slate of diverse designs based on the selected performance priorities. The optimal performance of the generated designs and the sensitivity of design performance to selected priorities are numerically confirmed. The present study demon strates both PIML -based and many -objective GD, which could allow engineers to improve designs along multiple performance goals in a streamlined manner.

Competition history

• JSHS 2025 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.