Digitized material design and performance prediction driven by high-throughput computing

Abstract

Introduction: The advancement of digitized material design has revolutionized the field of materials science by integrating computational modeling, machine learning, and high-throughput simulations. Traditional material discovery heavily relies on iterative physical experiments, which are often resource-intensive and time-consuming. Recent developments in high-throughput computing offer an efficient alternative by enabling large-scale simulations and data-driven predictions of material properties. However, conventional predictive models frequently suffer from limited generalization, inadequate incorporation of domain knowledge, and inefficient optimization of material structures. Methods: To address these limitations, we propose a novel framework that combines physics-informed machine learning with generative optimization for material design and performance prediction. Our approach consists of three major components: a graph-embedded material property prediction model that integrates multi-modal data for structure–property mapping, a generative model for structure exploration using reinforcement learning, and a physics-guided constraint mechanism that ensures realistic and reliable material designs. Results: By embedding domain-specific priors into a deep learning framework, our method significantly improves prediction accuracy while maintaining physical interpretability. Extensive experiments demonstrate that our approach outperforms state-of-the-art models in both predictive performance and optimization efficiency. Discussion: These findings highlight the potential of digitized design methodologies to accelerate the discovery of novel materials with desired properties and to drive next-generation material innovation.