OPMD: Phase-Field ML Benchmark

This OPMD (Open Phase-field Microstructure Dataset) provides a comprehensive benchmarking resource for evaluating machine learning strategies in the context of microstructure evolution governed by phase-field simulations. Derived as part of the study “Benchmarking machine learning strategies for phase-field problems” (Dingreville et al., 2024), this dataset includes high-resolution spatiotemporal fields generated using well-established phase-field models under various initial and boundary conditions. It features different microstructural morphologies, kinetics, and simulation parameters, allowing fair comparison of machine learning models across tasks such as temporal prediction, super-resolution, and microstructure encoding. This benchmark supports reproducible research in physics-informed machine learning and provides a foundation for developing surrogate models, uncertainty quantification methods, and hybrid physics–ML frameworks. If you use this dataset, please cite: R. Dingreville, A.E. Robertson, V. Attari, M. Greenwood, N. Ofori-Opoku, M. Ramesh, P.W. Voorhees, Q. Zhang. Benchmarking machine learning strategies for phase-field problems, Modelling Simul. Mater. Sci. Eng., 32 (2024) 065019. https://doi.org/10.1088/1361-651X/ad5f4a Hosted on Microstructures.net Microstructures.net is an open-access platform dedicated to the storage, visualization, and dissemination of microstructure datasets for materials science applications. The platform supports interactive exploration of image-based and field-based microstructure data and fosters community engagement in data-driven materials research. The OPMD dataset is hosted on Microstructures.net to facilitate broad accessibility and seamless integration with web-based tools for machine learning and visualization. Data Generation Overview To generate the microstructure dataset, we sampled 1,000 unique parameter combinations of [c^*, W, \kappa, M] using a Latin Hypercube Sampling (LHS) strategy. Unlike standard random sampling, LHS systematically divides each parameter’s range into equally probable intervals, ensuring uniform coverage of the multidimensional parameter space. Each simulation was conducted using the phase-field model described in the main text and evolved from time t = 0 to t = 10,000 (model units), with outputs recorded at 201 time steps. Each simulation trajectory captures the full microstructure evolution as a tensor of shape (200, 384, 384), representing 200 frames of spatial microstructure data over time. Unless otherwise specified, 700 simulations were used for training, 200 for testing, and 100 for validation in the corresponding paper.