A parametrically-Conditioned Deep Learning Surrogate for Coherent Spinodal Decomposition

A machine-learning framework is designed to tackle the spinodal decomposition of coherently strained alloy microstructures, with cubic anisotropy in elastic constants. A database of phase-field simulations, leveraging a Green's function approximation of the elastic field, is constructed for randomly chosen composition fields and widely variable misfit strain, producing a continuum variation in the phase morphology from smooth to strongly anisotropic domains. A convolutional recurrent neural network is then trained to accurately predict the full time-evolution sequence under the explicit conditioning of the known misfit parameter, at a reduced computational cost. Extensive error analysis at the pixel level and for global descriptors is used to assess the model accuracy and evaluate its generalization capability on longer timescales and larger computational domains. The model returns a one-to-one match of the ground-truth simulations over the temporal range of training sequences. Moreover, it can reliably predict average behaviors for sequences several times longer the training ones, albeit losing one-to-one accordance. As a proof, the NN trained model is used for reconstructing the full phase diagram of the system, achieving a degree of accuracy. The proposed framework is general and can be applied beyond the specific, prototypical system here considered, enabling high-throughput parametric studies.