Towards invertible 2D crystal structure representation for efficient downstream task execution

Abstract In the study of MoS 2 lattice defects, we explore the use of Siamese neural networks to create invariant embeddings, which respect the crystalline symmetry of the lattice. By training our model with contrastive learning, we successfully differentiate configurations with varying defects, achieving perfect accuracy in recognizing equivalent placements. Our method showcases the capability to predict physical properties like formation energy per site and the bandgap with strong performance across both low and high-defect density scenarios, outperforming traditional methods when enhanced with polynomial features. Despite its effectiveness, the model presents limitations at high defect densities, indicating a need for further refinement. Our approach lays the groundwork for reverse-engineering processes. Thus, we open pathways for generative models that can navigate from specified property ranges to optimal defect configurations, fostering an efficient solution-space exploration for bespoke material synthesis.