Research map / Image generation / Fast diffusion sampling
AdsorbDiff: Adsorbate Placement via Conditional Denoising Diffusion
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 AdsorbDiff, an SE(3)/T(2)-based conditional denoising diffusion model that predicts the lowest-energy adsorbate site and orientation on a slab, evaluated end-to-end with MLFF relaxation and DFT single points.
Problem
Finding the lowest-energy adsorbate configuration on a catalyst slab requires a non-convex global search over sites and orientations; prior approaches rely on heuristics or brute-force placements that are costly.
Approach
Formulate adsorbate placement as conditional denoising diffusion over 2D periodic translations and 3D rigid rotations of the adsorbate, conditioning the score on relative adsorption energies so that all ~100 DFT-optimized sites per adslab contribute training data rather than only the lowest-energy one.
Key results
- AdsorbDiff with a single site achieves a 31.8% DFT success rate versus 9.1% for AdsorbML with one site (about 3.5x).
- Pretraining on OC20 IS2RE local optima yields a zero-shot DFT success rate of 29.55%, and conditional training is reported to improve over unconditional training (11.36%).
- On OOD splits, AdsorbDiff drops only 3.8% relative to ID while still beating the AdsorbML single-site baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Condition diffusion training on relative energies instead of training unconditionally (ablation) | AdsorbDiff · OC20-Dense ID val | DFT success rate (%): 11.36 -> 31.81 | Helped |
| Predict sites with conditional diffusion rather than random single-site placement | AdsorbDiff vs AdsorbML · OC20-Dense ID val | DFT success rate (%): 9.1% -> 31.8% | Helped |
| Pretrain on OC20 local optimization data and evaluate zero-shot (ablation) | AdsorbDiff · OC20-Dense ID val | DFT success rate (%): 11.36 -> 29.55 | Helped |
| Conditionally fine-tune the pretrained diffusion model on OC20-Dense (ablation) | AdsorbDiff · OC20-Dense ID val | DFT success rate (%): 29.55 -> 31.81 | No clear effect |
| Use PaiNN instead of EquiformerV2 for diffusion (ablation) | AdsorbDiff · OC20-Dense ID val | DFT success rate (%): 27.3 vs 29.6 | No clear effect |
| Use GemNet instead of EquiformerV2 for diffusion (ablation) | AdsorbDiff · OC20-Dense ID val | DFT success rate (%): 27.3 vs 29.6 | No clear effect |
| Use ODE sampling instead of Annealed Langevin Sampling (ablation) | AdsorbDiff · Sampling and generalization beyond training data | sampling steps to convergence: 10x more steps | Helped |
Surprising findings
- AdsorbML outperforms AdsorbDiff at higher numbers of sites because brute-force random placements produce fewer anomalies.
- GNN architecture (PaiNN, GemNet-OC, EquiformerV2) has only marginal effect on diffusion success, unlike on force prediction.
Limitations
- Anomalies (desorption, dissociation, reconstruction, intercalation) reduce success rates, especially with multiple site predictions.
- Torsion angles are ignored, limiting applicability to larger adsorbates.
- Pretrained model learns local optima and may not capture global optima.
Experimental setup
Datasets: OC20 OC20-Dense
Benchmarks: OC20-Dense ID validation split OC20-Dense OOD validation split (50 of 200 samples)
Metrics: DFT success rate (within 0.1 eV of OC20-Dense benchmark) ML success rate anomaly rate
Baselines: AdsorbML Unconditional diffusion Random placement baseline
Builds on: DiffDock AdsorbML OC20 denoising score matching EquiformerV2
Closest papers by meaning
- DR-Label: Label Deconstruction and Reconstruction of GNN Models for Catalysis Systems (AAAI 2024 · similarity 0.84)
- Adjoint Sampling: Highly Scalable Diffusion Samplers via Adjoint Matching (ICML 2025 · similarity 0.82)
- Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 2024 · similarity 0.82)
- MOFDiff: Coarse-grained Diffusion for Metal-Organic Framework Design (ICLR 2024 · similarity 0.81)
- On scalable and efficient training of diffusion samplers (NeurIPS 2025 · similarity 0.81)
- Masked Diffusion Models as Energy Minimization (NeurIPS 2025 · similarity 0.81)
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.