Physics-informed machine learning for mapping of composition and mechanical properties in 316L/AlCrFeNi multi-material alloys fabricated by directed energy deposition

Abstract

The rapid design and property prediction of additively manufactured (AM) alloys are hindered by high experimental costs, data scarcity, and complex mechanical behaviours. To overcome these issues, a Gaussian process regression (GPR) framework integrating physical priors with AM experimental data (PS-GPR) was established to achieve efficient mapping among composition, microstructure, and mechanical properties. Utilising 316L/AlCrFeNi alloys as a model system, a physics-based model (PS model) derived from classical strengthening theories was first established as the baseline predictor. Subsequently, thirteen alloys with different mixing ratios were fabricated via AM, and their mechanical properties were experimentally measured. The residuals between the PS model predictions and experimental results were then learned and corrected using GPR, forming a physics-data hybrid gray-box modelling framework. The results demonstrate that the PS-GPR framework delivers superior predictive accuracy and stability across different training set sizes, attaining a minimum prediction error of 6.64%. Notably, this represents a 41.5% reduction in prediction error compared to a purely data-driven GPR model. The incorporation of physical models significantly improves prediction accuracy and robustness under limited sample conditions, providing an efficient and cost-effective pathway for rapid composition design and performance optimisation of new AM alloys.