Fig. 1: The active learning workflow.
From: Active learning of ternary alloy structures and energies
First, binary crystal configurations are sampled from the edges of the ternary phase diagram and input to the DFT code to evaluate formation energies. The dGCN (central block) is trained on these binary alloy formation energies, converting an input crystal into a graph, on which convolution and pooling operations are performed to convert it to a 42-element embedding vector representing the crystal. This vector is then fed to a feed-forward neural network with dropout to predict the formation energy and corresponding uncertainty. In the physics-informed scheme, ternary configurations are sampled from a particular composition for which the acquisition function is minimized and input to the DFT code for evaluation. In the data-driven scheme, the 42-element embedding vector is input to the DMaps algorithm (lower block) to discover latent coordinates. The acquisition function is computed on clusters formed in this lower dimensional space through k-means clustering, and ternary configurations are sampled from the selected cluster and input to the DFT code for evaluation. The dGCN is then retrained on a dataset containing these additional sampled ternary crystals.
