TrussGPT: large language model-driven inverse design framework for truss metamaterials

Mechanical metamaterials derive their exceptional properties from architected microstructures, yet inverse design for tailored performance remains computationally intensive and unintuitive. Here, we introduce TrussGPT, a large language model (LLM)-enabled generative framework that integrates natural language understanding with structural generation for automated inverse design of truss metamaterials. The framework comprises three key components. First, a bidirectional tokenizer encodes truss topologies into LLM-compatible sequences and decodes structural geometries from generated tokens. Second, a Transformer-based generative model constructs valid truss structures conditioned on specified mechanical targets. Third, a cross-domain fusion mechanism bridges semantic prompts and structural outputs, enabling intuitive, language-driven design. The framework supports a range of design scenarios, including single-property targeting, multi-property optimization, and complex stress–strain curve conditioning. TrussGPT achieves over 99% structural validity, maintains errors below 5% across targets, and generates designs up to 1000× faster than gradient-based or heuristic generative methods. Experimental results confirm the inverse design accuracy of TrussGPT-generated structures under specified constraints. This work establishes a scalable and interpretable paradigm for language-driven architected material design, with broad implications for intelligent manufacturing, aerospace engineering, and data-centric materials discovery.