Deep Learning for Predicting the Formation Energy and Synthesizability of Crystalline Materials

Predicting the stability and properties of hypothetical crystal structures is critical to accelerating materials discovery. This work introduces a neural network model that harnesses deep convolutional neural networks (CNNs) to analyze crystal voxel representations (CVRs) and predict a crystal's synthesizability and formation energy. CVRs represent crystals as 3D images, allowing CNNs to extract intricate structural motifs and make detailed predictions. Advanced CNN architectures, such as those with skip-connections, further enhance the model's capabilities. Moreover, rotational data augmentation enables the CNN to overcome limitations in handling symmetries and helps in the generalization of the predictions. The model achieves high performance, proficiently differentiating synthesizable from anomalous crystals and achieving a 0.046 eV/atom error for formation energy, on par with the state-of-the-art machine-learned predictive models for formation energy. Overall, this work demonstrates image representation learning for materials in computational materials science, where leading-edge CNNs analyze CVRs to enable precise crystal structure predictions and accelerate the discovery of novel materials.