Crys-JEPA: Accelerating Crystal Discovery via Embedding Screening and Generative Refinement

2026-05-14

AI Summary

Existing crystal generation methods struggle to simultaneously achieve structural stability and novelty. This work proposes Crys-JEPA, the first approach to adapt the Joint-Embedding Predictive Architecture (JEPA) to crystal generation. By constructing an energy-aware latent space that preserves formation energy differences, the method reformulates stability assessment as relative comparisons in the embedding space—eliminating reliance on costly energy computations or external references. Coupled with an embedding-based filtering strategy and an iterative retraining mechanism, Crys-JEPA substantially enhances generation quality. On the MP-20 and Alex-MP-20 benchmarks, it improves the V.S.U.N. metric by 81.4% and 82.6%, respectively, significantly outperforming current state-of-the-art baselines.

Abstract

De novo crystal generation seeks to discover materials that are not merely realistic, but also stable and novel. However, most existing generative models are trained to maximize the likelihood of observed crystals, which encourages samples to stay close to known materials yet not necessarily align with the criteria that matter in discovery. Through an empirical investigation, we show that current crystal generative models are caught in a pronounced stability--novelty trade-off: moving toward the observed distribution preserves stability but limits novelty, whereas moving away from it quickly destroys stability. This suggests that the useful region for discovering crystals that are both stable and novel is extremely narrow. To escape the trade-off, we introduce Crys-JEPA, a joint embedding predictive architecture for crystals that learns an energy-aware latent space preserving formation-energy differences. In this space, stability assessment can be reformulated as an embedding-based comparison against accessible training crystals, reducing the reliance on expensive energy evaluation and task-specific external references. Building on Crys-JEPA, we further develop a screening-and-refinement pipeline that identifies promising generated crystals and reintroduces them to refine the generative model. On MP-20 and Alex-MP-20 datasets, we achieve improvements over baselines up to 81.4% and 82.6% on V.S.U.N metric, respectively.

Problem

• crystal generation
• stability–novelty trade-off
• de novo materials discovery
• formation energy
• generative models

Innovation

• Crys-JEPA
• energy-aware latent space
• stability-novelty trade-off
• embedding-based screening
• generative refinement