Answers are grounded in this paper and cite their sources.
The combination of artificial intelligence and materials science promises accelerated discovery, yet existing approaches remain limited by a fundamental divide between physically grounded prediction and flexible reasoning. Large atomic models provide accurate structure-property relationships but are computationally expensive and lack general reasoning capability, whereas large language models excel at abstraction and inference but struggle with structured physical representations and reliable quantitative prediction. Efforts to bridge these paradigms have been impeded by competing training objectives: atomic models demand precise numerical optimization, while reasoning requires coherent sequential inference; when forced into a single objective, these requirements interfere, degrading both predictive fidelity and logical consistency. Here we introduce a Large Language and Atomic Model (LLAM), a multimodal framework with native structural reasoning capability that directly integrates atomistic representations with language-based inference. By aligning atomistic representations with the language embedding space through a lightweight cross-modal bridge, LLAM conditions both quantitative prediction and subsequent reasoning directly on crystal structures. Across diverse benchmarks spanning continuous property prediction and crystallographic classification, LLAM demonstrates strong performance while exhibiting improved reasoning capability. Case studies in optoelectronic-property interpretation, solid-state electrolyte optimization, and photovoltaic-material screening further demonstrate that LLAM can connect structure-grounded predictions with physically meaningful interpretations and application-oriented decisions.
No takes yet. Share an insight, caveat, or question.
He et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: