Mechanical Property Prediction of Laser-Processed H13 Tool Steel Informed by Phase Composition and Grain Morphology via Machine Learning

Abstract Process-structure-property relationships are highly nonlinear, multiscale, and difficult to capture with traditional analytical models. Yet tools that harness these relationships can address the increasingly demanding needs of modern manufacturing, where rapid and reliable assessment of material performance is essential. This potential has intensified interest in data-driven approaches capable of inferring mechanical behavior without protracted inspection or costly destructive testing. Here, machine-learning models are used to predict the mechanical properties of laser-processed H13 tool steel by drawing on microstructure-property relationships. Firstly, published results and data from samples produced by Directed Energy Deposition were compiled into datasets. Phase recognition capabilities were then developed through a deep learning, Convolutional Neural Network scheme that was applied to obtain the phase composition of each sample. Due to their high strength and hardness, the phase fraction of martensite or tempered martensite was used as an indicator of mechanical performance. Grain morphology data were also extracted from processed microstructural images. Finally, a custom Artificial Neural Network, designed to balance the influence of phase fraction and grain morphology, was trained and validated for mechanical property prediction. Correlation coefficients exceeding R 2 = 0.939 were achieved for Ultimate Tensile Strength, Yield Strength, and Elongation, with root mean square testing errors of 36, 42 MPa, and 0.6%, respectively. These results demonstrate the predictive power of machine learning-driven microstructural analysis and its potential to complement existing inspection techniques.