AdsorbDiff: Adsorbate Placement via Conditional Denoising Diffusion

TL;DR

AdsorbDiff reframes adsorbate placement on slabs as a conditional denoising diffusion problem that respects the translation $T(2)$ and rotation $SO(3)$ symmetries and slab periodicity. By conditioning the diffusion process on relative energies $E_{ ext{rel}}^c$ and integrating with a pretrained ML force field followed by DFT validation, the approach significantly improves the search for the lowest-energy adslab configuration, achieving a DFT success rate of $E_{Ads}^{DFT}$ within 0.1 eV in 31.8% of trials with a single site prediction—roughly 3.5x better than the naive AdsorbML baseline. The study demonstrates that conditional diffusion coupled with end-to-end evaluation yields substantial speedups and is robust to GNN architectural choices, while benefiting from pretraining on large OC20 local optima and generalizing to unseen adsorbates and slabs, with manageable inference costs. This work offers a scalable pathway to accelerate catalyst discovery by reducing reliance on brute-force site sampling and expensive DFT checks, enabling more extensive exploration of adsorbate–slab configurations. It also highlights areas for improvement, such as handling anomalies, incorporating torsion angles for larger adsorbates, and refining constraint strategies for larger systems.

Abstract

Determining the optimal configuration of adsorbates on a slab (adslab) is pivotal in the exploration of novel catalysts across diverse applications. Traditionally, the quest for the lowest energy adslab configuration involves placing the adsorbate onto the slab followed by an optimization process. Prior methodologies have relied on heuristics, problem-specific intuitions, or brute-force approaches to guide adsorbate placement. In this work, we propose a novel framework for adsorbate placement using denoising diffusion. The model is designed to predict the optimal adsorbate site and orientation corresponding to the lowest energy configuration. Further, we have an end-to-end evaluation framework where diffusion-predicted adslab configuration is optimized with a pretrained machine learning force field and finally evaluated with Density Functional Theory (DFT). Our findings demonstrate an acceleration of up to 5x or 3.5x improvement in accuracy compared to the previous best approach. Given the novelty of this framework and application, we provide insights into the impact of pre-training, model architectures, and conduct extensive experiments to underscore the significance of this approach.

Figures (10)

• Figure 1: Overview of AdsorbDiff: Random initial site and orientation for the adsorbate are selected, followed by sampling over 2D translation, 3D rigid rotations, and considering periodic boundary conditions (PBC) to predict the optimal site and orientation. MLFF optimization is then conducted from the predicted site with a fixed interstitial gap until convergence. The final prediction undergoes constraint verification, and DFT verification is performed on valid structures to calculate success rates.
• Figure 2: Comparison of conditional and unconditional diffusion with a baseline of random placement. Conditional diffusion training on relative energies of configurations of adslab significantly improves success rates over unconditional training and AdsorbML baseline.
• Figure 3: DFT Success Rates (%) for AdsorbDiff and AdsorbML across a varying number of site predictions. AdsorbDiff performs 3.5x better than AdsorbML utilizing a single site prediction. At higher sites, AdsorbML performs better due to the brute-force nature of site prediction that reduces anomalies.
• Figure 4: Anomalies in AdsorbDiff and AdsorbML with respect to Nsites. A system is labeled as anomalous if all its predicted sites result in anomalies. AdsorbML has fewer anomalies than AdsorbDiff at higher Nsites due to more randomness in initial sites.
• Figure 5: Impact of pretraining on 460k OC20 local optima data on DFT Success Rate. PT Zero-shot measures zero-shot generalization of OC20 pre-trained model to OC20-Dense data. PT Conditional is finetuned on OC20 Dense data conditionally on relative energies of adslab configurations. Random baseline corresponds to randomly placed adsorbate.