Integrating 3D CNN and phase-field simulation for TiAlN coating property prediction via 3D microstructure

Microstructure evolution in service significantly influences the properties of advanced materials. Numerical simulation can effectively capture microstructure development and obtain abundant high-fidelity data. However, there is a lack of effective 3D microstructure-informed property prediction methods due to the complexity and informativeness of 3D microstructural data. In this work, a novel approach combining phase-filed simulation and 3D convolutional neural network was proposed to explore the composition-process-structure-property relation in Ti_1-xAl_xN coatings. A large dataset of 4,962 simulated 3D microstructures under various heat treatment conditions was first generated using phase-field simulations. Then, a reconstructible feature extraction model was trained to compress each 48 × 48 × 48 grids microstructure into a 128-dimensional latent vector with a reconstruction accuracy of up to 99%. Using the extracted features, a microstructure-based hardness prediction model was constructed, reaching a low prediction error of 1.6 GPa (ca. 5.3% error for an average hardness of 30.8 GPa). The results demonstrate the effectiveness of 3D microstructure-informed deep learning method for accurate property prediction, providing a promising tool for the data-driven design of high-performance materials.