AdsorbML: A Leap in Efficiency for Adsorption Energy Calculations using Generalizable Machine Learning Potentials

Abstract

Computational catalysis is playing an increasingly significant role in the design of catalysts across a wide range of applications. A common task for many computational methods is the need to accurately compute the minimum binding energy — the adsorption energy — for an adsorbate and a catalyst surface of interest. Traditionally, the identification of low energy adsorbate-surface configurations relies on heuristic methods and researcher intuition. As the desire to perform high-throughput screening increases, it becomes challenging to use heuristics and intuition alone. In this paper, we demonstrate machine learning potentials can be leveraged to more accurately and efficiently identify low energy adsorbate-surface configurations. Our DFT verified algorithm provides a spectrum of performance trade-offs, with one balanced option finding the lowest energy configuration, within a 0.1 eV threshold, 86.33% of the time, while achieving a 1331x speedup in computation. To standardize benchmarking, we introduce the Open Catalyst Dense dataset containing nearly 1,000 diverse surfaces and 85,658 unique configurations.

Beyond Independent Error Assumptions in Large GNN Atomistic Models

“…This dataset was created to address the issue that the the adsorption energies in the OC20 dataset 22 may not correspond to the minimum energy of a specific adsorbate-catalyst combination. The OC20-Dense dataset collects 995 unique adsorbate-catalyst combinations from the OC20 dataset and densely enumerates initial configurations of adsorbates on surfaces, using both 'heuristic' and 'random' strategies 8 . The heuristic strategy utilizes popular tools like CatKit 31 and Pymatgen 32 , while the random strategy places the adsorbate on randomly selected sites on the surface.…”

“…In this research, the relaxation results on the configurations established by the heuristic approach are used as a baseline dataset since the DFT relaxations on them represent the common community baseline for evaluating multiple adsorption energies on the surfaces 31,32 . OC20-Dense-Heuristic comprises 995 unique adsorbate-catalyst combinations spanning 76 adsorbates and 850 bulk catalysts, with an average of 31.2 configurations per combination 8 . There are a total of 31,081 adsorbate-catalyst systems, resulting in approximately 483 million possible system pairs.…”

“…In practical applications, the relative difference in adsorption energies (∆∆E ads ) between chemically similar systems is of particular interest, as described in Figure 1. For example, the adsorption energy of a specific surface can be obtained by identifying the minimum adsorption energy on the surface 7,8 . In this case, the adsorption energies of different configurations should be compared.…”

Deep Learning-Based Energy Mapping of Chlorine Effects in an Epoxidation Reaction Catalyzed by a Silver–Copper Oxide Nanocatalyst

“…Moreover, the efficiency of graph neural networks (GNNs) in determining relative energy differences has been scrutinized, revealing nuanced insights into error assumptions . Augmenting these foundational developments, machine learning has been adeptly used to expedite adsorption energy calculations, showcasing remarkable computational efficiency gains. , …”

“…38 Augmenting these foundational developments, machine learning has been adeptly used to expedite adsorption energy calculations, showcasing remarkable computational efficiency gains. 39,40 However, within the scope of our study, we adopt a more streamlined and targeted approach. While the aforementioned literature delves into advanced techniques and novel potentials, our focus remains anchored on the foundational level: training the neural network based solely on atomic positions, charges, and their associated energies.…”