Research map / Image generation / Fast diffusion sampling
Chemistry-Inspired Diffusion with Non-Differentiable Guidance
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 ChemGuide, which uses the non-differentiable quantum chemistry oracle GFN2-xTB as guidance for an unconditional 3D latent diffusion model (GeoLDM) via zeroth-order SPSA gradient estimates, reducing atomic forces and combining with neural guidance.
Problem
Guided molecular diffusion normally needs labeled data to train property predictors or conditional models, which is costly and generalizes poorly; generated 3D molecules also often have high atomic forces (unstable geometries).
Approach
Treat a non-differentiable quantum chemistry oracle (GFN2-xTB) as guidance for an unconditional latent diffusion model, estimating its gradient with zeroth-order SPSA perturbations of atom coordinates that preserve zero center of gravity; combine with neural guidance in a bilevel optimization.
Key results
- On QM9 with GeoLDM and xTB guidance at scale 0.0001, force RMS dropped 6.76% (0.0104 vs 0.0111) and validity rose 1.60% (91.40% vs 89.80%).
- On GEOM at guidance scale 0.1, force RMS fell 14.16% (0.0411 vs 0.0478) and energy above ground state fell 13.92%.
- With noisy neural guidance on unconditional GeoLDM, MAE improved over unconditional GeoLDM by 57.92% (alpha) and 77.09% (Cv); clean guidance beat C-GeoLDM on mu by 81.68%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GFN2-xTB force guidance to GeoLDM | GeoLDM · QM9 | Force RMS (Eh/Bohr): 0.0111 -> 0.0104 | Helped |
| Add GFN2-xTB force guidance to GeoLDM | GeoLDM · QM9 | Energy above Ground State (Eh): 0.0050 -> 0.0042 | Helped |
| Add GFN2-xTB force guidance to GeoLDM | GeoLDM · GEOM | Force RMS (Eh/Bohr): 0.0478 -> 0.0411 | Helped |
| Add noisy neural guidance | GeoLDM · QM9, static polarizability alpha | MAE: 5.6732 -> 2.3870 | Helped |
| Add clean neural guidance | GeoLDM · QM9, dipole moment mu | MAE: 1.6461 -> 0.2031 | Helped |
| Add clean neural guidance | GeoLDM · QM9, static polarizability alpha | MAE: 5.6732 -> 5.4798 | No clear effect |
| Combine explicit property guidance with ChemGuide stability guidance | C-GeoLDM · QM9, static polarizability alpha | MAE: 2.5593 -> 2.4430 | Helped |
| Combine explicit property guidance with ChemGuide stability guidance | C-GeoLDM · QM9, Delta epsilon | MAE: 0.4982 -> 0.5225 | Hurt |
| Combine noisy neural guidance with ChemGuide stability guidance | C-GeoLDM · QM9, dipole moment mu | MAE: 1.1582 -> 0.7246 | Helped |
| Use xTB oracle guidance to optimize molecular properties | GeoLDM · QM9, epsilon_HOMO | MAE: 0.6264 -> 0.6392 | Hurt |
| Use xTB oracle guidance to optimize molecular properties | GeoLDM · QM9, dipole moment mu | MAE: 1.5490 -> 1.5295 | Helped |
| Use xTB oracle guidance to optimize molecular properties | GeoLDM · GEOM, epsilon_HOMO | MAE: 1.1922 -> 1.2939 | Hurt |
Surprising findings
- Unconditional GeoLDM with clean neural guidance outperforms conditionally trained C-GeoLDM on several properties (mu, Cv, HOMO, Delta-epsilon).
- Gains from ChemGuide on GEOM (larger molecules) were larger than on QM9 for force reduction.
Limitations
- xTB runs on CPU, making guided sampling slow and GPU-inefficient
- Guidance scale is hard to choose in advance and performance is sensitive to it
- Improvements in validity and stability are modest (about 1-2 points)
Experimental setup
Datasets: QM9 GEOM
Benchmarks: QM9 GEOM
Metrics: force RMS validity uniqueness atom stability molecule stability energy above ground state MAE
Baselines: EDM GeoLDM C-EDM C-GeoLDM noisy neural guidance clean neural guidance
Builds on: GeoLDM EDM classifier guidance SPSA GFN2-xTB
Closest papers by meaning
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials (ICML 2025 · similarity 0.89)
- Geometric-Facilitated Denoising Diffusion Model for 3D Molecule Generation (AAAI 2024 · similarity 0.88)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.88)
- Graph Diffusion Transformers for Multi-Conditional Molecular Generation (NeurIPS 2024 · similarity 0.87)
- DiSCO: Diffusion Schrödinger Bridge for Molecular Conformer Optimization (AAAI 2024 · similarity 0.87)
- Conditional Synthesis of 3D Molecules with Time Correction Sampler (NeurIPS 2024 · similarity 0.87)
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.