SLICES-PLUS

In recent years, the realm of crystalline materials has witnessed a surge in the development of generative models, predominantly aimed at the inverse design of crystals with tailored physical properties. However, spatial symmetry, which serves as a significant inductive bias, is often not optimally harnessed in the design process. This oversight tends to result in crystals with lower symmetry, potentially limiting the practical applications of certain functional materials. To bridge this gap, we introduce SLICES-PLUS, an enhanced variant of SLICES that emphasizes spatial symmetry. Our experiments in classification and generation have shown that SLICES-PLUS exhibits greater sensitivity and robustness in learning crystal symmetries compared to the original SLICES. Furthermore, by integrating SLICES-PLUS with a customized MatterGPT model, we have demonstrated its exceptional capability to target the specific physical properties and crystal systems with precision. Finally, we explore autoregressive generation towards multiple elastic properties in few-shot learning. Our research represents a significant step forward in the realm of computational materials discovery.