Abstract <p>&#xa0;Understanding and predicting microstructure evolution is fundamental to materials science, as it governs the resulting properties and performance of materials. However, traditional phase-field simulations, while physically accurate, are computationally expensive because they require repeated solutions of nonlinear partial differential equations over fine spatiotemporal grids. In this work, we aim to develop an efficient and scalable surrogate framework for accelerated microstructure evolution prediction. To this end, we investigate a fully convolutional, nonrecurrent spatiotemporal learning model trained in a self-supervised manner on sequential phase-field simulation data for grain growth and spinodal decomposition, allowing the network to learn the underlying evolution dynamics directly from data. Numerical experiments demonstrate that the proposed framework accurately captures both short-term interfacial dynamics and long-term statistical characteristics of evolving microstructures. The model also generalizes from low-resolution training data to substantially higher-resolution simulations without architectural modification or retraining, demonstrating strong resolution scalability. In addition, runtime and FLOPs analyses indicate substantially lower inference cost than recurrent spatiotemporal architectures, supporting efficient long-horizon forecasting. These results indicate that fully convolutional spatiotemporal learning provides a computationally efficient and scalable complement to traditional PDE-based simulation and recurrent neural forecasting methods, establishing a practical pathway toward accelerated microstructure modeling and digital-twin-enabled materials design.</p> Graphical abstract <p></p>