PRISMat: Policy-Driven, Permutation-Invariant Autoregressive Material Generation
The paper introduces PRISMat, a new model for material generation that is both cost-effective and permutation-invariant. It aims to improve the efficiency of material discovery by outperforming large language models in generating crystal slabs based on surface properties. The authors report significant reductions in error rates for key material properties compared to existing models.
- ▪PRISMat is designed to address the inefficiencies of large language models in material generation.
- ▪The model achieves mean absolute errors of 0.188 eV/A$^2$ for cleavage energy and 2.79 eV for work function tasks.
- ▪PRISMat reduces the error of the next best model by four times.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.16612 (cs) [Submitted on 15 May 2026] Title:PRISMat: Policy-Driven, Permutation-Invariant Autoregressive Material Generation Authors:Claire Schlesinger, Circe Hsu, Peter Schindler, Robin Walters View a PDF of the paper titled PRISMat: Policy-Driven, Permutation-Invariant Autoregressive Material Generation, by Claire Schlesinger and 3 other authors View PDF HTML (experimental) Abstract:Rapid identification of candidate materials with target properties has become a key task in materials science.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
