Research map / Probabilistic and scientific modeling / Molecules and proteins

CrystalREPA: Transferring Physical Priors from Universal MLIPs to Crystal Generative Models

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine-learned interatomic potentials · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CrystalREPA, which aligns atom-wise hidden states of crystal generative models with frozen universal MLIP representations via an element-aware contrastive loss, improving stability and validity across three generators and ten teachers.

Problem

Crystal generative models learn what stable crystals look like but receive little supervision about what makes them stable (potential energy surface), leaving a representation gap relative to pretrained universal MLIPs.

Approach

Align atom-wise hidden states of a generative encoder (on the corrupted structure) with frozen MLIP atom-wise representations of the clean crystal using an element-aware InfoNCE that removes same-element negatives; the projection head and teacher are discarded at inference. Teacher transfer effectiveness is predicted by representation distinguishability (CRI/ERI/TRI) rather than Matbench Discovery

Key results

• On MP-20, averaged over 10 MLIP teachers, CrystalREPA raises Metastable% from 41.2 to 50.5 for CrystalFlow, 57.4 to 62.7 for MatterGen, and 43.2 to 50.9 for DiffCSP.
• On Alex-MP-20 with DPA-3.1-3M teacher, MSUN% improves from 31.5 to 33.7 (CrystalFlow), 38.1 to 39.3 (MatterGen), and 35.0 to 36.3 (DiffCSP).
• Metastable rate correlates with teacher CRI, ERI, and TRI at 0.863, 0.857, and 0.868 Pearson, while Matbench Discovery CPS shows -0.120.

Surprising findings

• Teacher leaderboard accuracy (Matbench Discovery CPS) is not predictive of transfer effectiveness; SevenNet-Omni-i12, the top CPS model, gives relatively small gains.
• Raw Novel% decreases in some MP-20 settings even as MSUN% and SUN% improve.

Limitations

• Aligns only atom-level representations; no edge-level or higher-order structural alignment.
• Evaluated only on crystalline materials; transfer to molecules or proteins untested.

Experimental setup

Datasets: MP-20 Alex-MP-20

Benchmarks: MP-20 Alex-MP-20 Matbench Discovery

Metrics: Metastable% MSUN% Stable% SUN% E_hull RMSD Valid% Unique% Novel% formation energy probing MAE

Baselines: Vanilla CrystalFlow Vanilla MatterGen Vanilla DiffCSP standard InfoNCE cosine alignment energy-conditioned generation

Builds on: REPA (representation alignment for diffusion) InfoNCE CrystalFlow MatterGen DiffCSP

Closest papers by meaning

• Scalable Diffusion for Materials Generation (ICLR 2024 · similarity 0.89)
• WyckoffDiff -- A Generative Diffusion Model for Crystal Symmetry (ICML 2025 · similarity 0.87)
• All that structure matches does not glitter (NeurIPS 2025 · similarity 0.87)
• LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 2025 · similarity 0.87)
• Crystalformer: Infinitely Connected Attention for Periodic Structure Encoding (ICLR 2024 · similarity 0.87)
• Wyckoff Transformer: Generation of Symmetric Crystals (ICML 2025 · similarity 0.86)