Journal DOI 10.68956/journal.human-performance-futures

Creative Commons source reading · Machine learning for performance

A Closed‐Loop Framework for Inverse Design: Dynamic Training and Intelligent Optimization for Heterostructured Materials

Zhiyan Zhong; Xueru Zheng; Xiao Zhou; Zhengheng Tao; Wan Han; Jingyi Pan; Dian Wu; Ning Gao; Lei Liu; Zhongyang Wang; Fanchao Meng; Tongxiang Fan

Originally published in Adv Sci (Weinh) · 2026-07-20

From the article

Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.

Abstract

ABSTRACTFor decades, heterostructured materials have attracted great research interest because of their remarkable mechanical and physical properties. Nevertheless, exploring the immense design space to address the intrinsic trade‐off between strength and toughness through trial‐and‐error experiments and simulation approaches entails considerable temporal and economic expenditures. Here, we propose a comprehensive end‐to‐end scientific machine learning framework integrating a deep learning model (Back‐Propagation Neural Network with Continual Learning, BPNN‐CL) for forward prediction and an optimization algorithm (NSGA‐II with Partition Monitoring and Chaotic Perturbation, NSGA‐II‐PMCP) for inverse optimization, and demonstrate its effectiveness using metal matrix composites as representative heterostructured materials. The forward…

Conclusions

In this study, we establish a state‐of‐the‐art framework integrating BPNN‐CL and NSGA‐II‐PMCP for the inverse design of different HSMMCs. For prediction, R2 values for UTS and Kt exceed 0.98 and 0.94, respectively. Against seven mainstream static methods, BPNN‐CL is consistently the best. External validation on experiments shows most UTS and E predictions fall within ±10% of reported values, even for some absent from training, which demonstrates meaningful extrapolative power. For inverse exploration, the optimized UTS‐Kt Pareto frontiers for Al2014, Al6061, and Al7075‐based PRAMCs all perform better than Bayesian optimization. Specifically, Al7075‐based PRAMCs show 68.95% higher HV than standard NSGA‐II. Also, Al6061‐based PRAMCs designs achieve better UTS‐Kt with faster convergence than RFR‐NSGA‐II. Furthermore, the Al6061‐based…

Editorial reading guide

  • What data created and tested the model, and was evaluation independent?
  • What is the cost of a false positive, false negative or out-of-distribution decision?