A Decision Transformer Approach to Grain Boundary Network Optimization

TL;DR

This paper introduces a Decision Transformer ML model trained on human-in-the-loop optimization data to efficiently solve complex grain boundary network design problems, outperforming traditional algorithms in iteration efficiency.

Contribution

The study demonstrates how a Decision Transformer can learn from human optimization trajectories and generalize across different microstructure models, advancing materials design methods.

Findings

ML model achieves 84% validation accuracy against human decisions

Solutions comparable to simulated annealing with 1/1000th of the iterations

Model generalizes to different microstructure models without retraining

Abstract

As microstructure property models improve, additional information from crystallographic degrees of freedom and grain boundary networks (GBNs) can be included in microstructure design problems. However, the high dimensional nature of including this information precludes the use of many common optimization approaches and requires less efficient methods to generate quality designs. Previous work demonstrated that human-in-the-loop optimization, instantiated as a video game, achieved high-quality, efficient solutions to these design problems. However, such data is expensive to obtain. In the present work, we show how a Decision Transformer machine learning (ML) model can be used to learn from the optimization trajectories generated by human players, and subsequently solve materials design problems. We compare the ML optimization trajectories against players and a common global optimization algorithm: simulated annealing (SA). We find that the ML model exhibits a validation accuracy of 84% against player decisions, and achieves solutions of comparable quality to SA (92%), but does so using three orders of magnitude fewer iterations. We find that the ML model generalizes in important and surprising ways, including the ability to train using a simple constitutive structure-property model and then solve microstructure design problems for a different, higher-fidelity, constitutive structure-property model without any retraining. These results demonstrate the potential of Decision Transformer models for the solution of materials design problems.