Simulation-informed deep learning for estimating material parameters from microstructural morphology

Accurate estimation of material parameters is essential for predictive physics-based modeling of microstructure evolution, yet phase-field parameters such as gradient energy coefficients and effective mobilities are difficult to determine directly and often require computationally expensive calibration. In this work, we develop a simulation-informed deep learning framework for estimating phase-field parameters from microstructural morphology. The framework is demonstrated for spinodal decomposition in Fe–36 wt% Cr aged at 773 K for 100 h. Synthetic microstructures were generated using Cahn–Hilliard phase-field simulations over a predefined ( κ , M ) parameter space, where κ is the gradient energy coefficient and M is the effective mobility. A convolutional neural network with a frozen EfficientNetB7 feature extractor and multi-task regression head was trained to infer κ and M directly from microstructure images. The model showed strong predictive performance on held-out synthetic data and generalized to off-grid parameter combinations without retraining. Extended-mobility and morphology-based sensitivity analyses showed that prediction accuracy decreases when highly coarsened morphologies become similar across different parameter combinations. A proof-of-concept application to an experimental Fe–Cr microstructure produced parameter estimates consistent in order of magnitude with representative literature-scale values. Although demonstrated here for a specific Fe–Cr spinodal-decomposition case, this work shows that simulation-informed deep learning can provide an efficient pathway for repeated image-based parameter estimation when microstructural features remain sensitive to the governing parameters.