MatterGPT: A Generative Transformer for Multi-Property Inverse Design of Solid-State Materials

Abstract Inverse design of solid-state materials with desired properties remains a central challenge in materials science, requiring exploration of vast chemical spaces containing potentially 10¹⁰⁰ possible structures. Current generative approaches face limitations in computational efficiency, multi-property targeting precision and mechanistic interpretability. Here, we introduce MatterGPT, an autoregressive Transformer-decoder architecture that leverages SLICES (Simplified Line-Input Crystal-Encoding System) representation to generate novel crystals through conditional next-token prediction. Trained on 306,533 crystal structures, MatterGPT achieves > 99% structural validity, > 99% structural uniqueness and > 50% novelty rates while targeting both specific lattice-insensitive and lattice-sensitive properties. Critically, MatterGPT enables direct multi-property generation without post-generation filtering. Interpretability analysis reveals clear property-guided generation mechanisms and systematic chemical space exploration. The comprehensive open-source release, including MatterGPT Hub integration platform, establishes sequence-based autoregressive generation as a computationally efficient and interpretable paradigm for inverse crystal design, accelerating materials discovery across energy storage, electronics, and functional applications.