(532eb) Infusing Theory into Deep Learning for Interpretable Stability Prediction of Transition Metal Alloys

Transition metal alloys (TMA) play a significant role in heterogeneous catalysis due to their great tunability in surface adsorption properties. However, the bottleneck of TMA catalyst design often lies on the mismatch between desired reactivity and unsatisfying stability, which calls for effective TMA stability prediction over the huge chemical space. Although many advanced machine learning (ML) models have been reported to achieve promising accuracy on TMA formation energy prediction, the lack of interpretability stemmed from their black-box nature prevents us from extracting chemical insights from model predictions and also greatly hinders the improvement of model generalization for new materials. In the present work, we develop an interpretable deep learning algorithm for TMA formation energy prediction via infusing cohesion theory of TMA into graph-based convolutional neural networks (CNNs). Specifically, a CNN model is trained to learn a feature vector for each atomic site and a fully connected network is jointly trained to map the learned atomic feature to primary physical parameters of the four cohesive energy contribution terms, i.e., promotion, renormalization, conduction band formation and d-band formation. Finally, the cohesion theory is utilized to calculate the total formation energy based on these predicted physical parameters. Among the predicted outcomes, not only the total formation energy which characterizes the crystal stability can be obtained, but also different contribution terms and intuitive physical parameters such as the number of localized electrons which describe the underlying physics picture are revealed. Since the training process is guided by the physics principle, the model is believed to have lower risk of overfitting and superior generalization compared to purely data-driven ML models. Potential applications of this model include transfer learning between bulk and surface structures and transferable ML potentials for multi-principle-element TMA systems.