Research map / Probabilistic and scientific modeling / Molecules and proteins
Latent Diffusion Pretraining for Crystal Property Prediction
Shrimon Mukherjee, KISHALAY DAS, Partha Basuchowdhuri, Pawan Goyal, Niloy Ganguly
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add latent diffusion pretraining on top of PDDFormer encoder | PDDFormer · JARVIS-DFT Bulk Modulus | MAE: 9.546 -> 8.817 | Improved |
| add latent diffusion pretraining on top of Matformer encoder | Matformer · JARVIS-DFT Ehull | MAE: 0.064 -> 0.045 | Improved |
| do latent-space diffusion instead of feature-space diffusion (DPF) | crystal pretraining · JARVIS-DFT Bandgap (MBJ) | MAE: 0.311 -> 0.238 | Improved |
| add LDM stage after VAE pretraining (ablation) | CrysLDNet pretraining · JARVIS-DFT Bulk Modulus | MAE: 10.61 -> 8.817 | Improved |
| add VAE pretraining before latent diffusion (ablation) | CrysLDNet pretraining · JARVIS-DFT Bandgap (MBJ) | MAE: 0.302 -> 0.238 | Improved |
| LDM-only pretraining versus VAE-only pretraining (ablation) | CrysLDNet pretraining · JARVIS-DFT (4 properties) | MAE: Bulk 10.61 (VAE) vs 10.37 (LDM); MBJ 0.284 (VAE) vs 0.302 (LDM) | Mixed |
| reconstruct all of atom types, coordinates and lattice in the VAE instead of a subset (ablation) | CrysLDNet VAE · JARVIS-DFT SLME | MAE: 4.678 -> 4.120 | Improved |
| train LDM longer to refine VAE latents (5 to 50 epochs) (ablation) | CrysLDNet (Matformer) · JARVIS-DFT Bulk Modulus | MAE: 10.35 -> 9.818 | Improved |
| latent diffusion refinement raises mutual information of latents (ablation) | VAE-only encoder · GNoME pretraining set, MI with atom types | mutual information: 3.0906 -> 4.5465 | Improved |
| pretrain with CrysLDNet when finetuning data is scarce (20-40%) | Matformer and PDDFormer · JARVIS-DFT limited data | MAE: average improvement 12.83% and 22.49% at 40% data | Improved |
| pretrain with CrysLDNet when data is plentiful (large property sets) | PDDFormer · JARVIS-DFT large-data properties | MAE: average gain 3.0% (large data) vs 5.27% (small data) | Improved |
| pretrain with CrysLDNet then finetune on DFT, zero-shot on experiment | CrysGNN, CrysDiff, DPF · OQMD-EXP formation energy | MAE: 0.211 (CrysDiff) -> 0.205 | Improved |
Closest papers by meaning
- A Diffusion-Based Pre-training Framework for Crystal Property Prediction (AAAI 2024 · similarity 0.93)
- Vector Field Oriented Diffusion Model for Crystal Material Generation (AAAI 2024 · similarity 0.91)
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.90)
- Periodic Materials Generation using Text-Guided Joint Diffusion Model (ICLR 2025 · similarity 0.90)
- LLM Meets Diffusion: A Hybrid Framework for Crystal Material Generation (NeurIPS 2025 · similarity 0.89)
- A Denoising Pre-training Framework for Accelerating Novel Material Discovery (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
