Latest Artificial Intelligence (AI) Research From CMU And Meta Demonstrates How Machine Learning’s Potential Can Be Leveraged To Identify Low Energy Adsorbate-Surface Configurations More Accurately And Efficiently

Over the past decade, computational catalysis has emerged as one of the most active research areas, and it is currently an essential tool for studying chemical processes and active sites. The requirement to calculate exactly the lowest binding energy (the adsorption energy) for an adsorbate and a catalyst surface of interest is a typical challenge for many computational approaches. Heuristic techniques and researcher intuition have traditionally been used to identify low-energy adsorbate-surface combinations. Unfortunately, using heuristics and intuition alone becomes more difficult as the need for high-throughput screening increases.

It is necessary to relax the atomic positions until a local energy minimum is reached to calculate the adsorption energy for a particular adsorbate surface configuration. The most popular method for performing this adsorbate surface relaxation is Density Functional Theory (DFT). The required DFT calculations can take days or weeks as several configurations are generally explored to estimate the adsorption energy. Recent advances in the estimation of atomic force and energy using machine learning (ML) potentials, which are orders of magnitude faster than DFT, have been seen. In this context, a research team from Carnegie Mellon University proposed AdsorbML, a hybrid approach to estimate adsorption energies that uses the advantages of both ML potentials and DFT. The adsorbML algorithm uses ML to accelerate the adsorbate placement process and identify the adsorption energy under a spectrum of accuracy-efficiency trade-offs.

AdsorbML uses GPUs to perform ML relaxations and ranks them from lowest to maximum energy. The top k systems are sent to DFT for either a full relaxation of the ML relaxed structure (RX) or a single point evaluation (SP). Systems that do not meet the constraints are filtered out at each stage of the relaxation process. The minimum of all DFT values ​​is considered for the final adsorption energy.

In addition, the authors developed the Open Catalyst 2020 – Dense Dataset (OC20-Dense) to measure the task of adsorption energy search. By examining multiple configurations for each distinct adsorbate surface system, OC20-Dense approximates the actual adsorption energy. OC20-Dense consists of 87,045 randomly and heuristically generated configurations, 850 inorganic bulk crystal structures, and 995 different adsorbate-surface pairings spanning 76 different adsorbates. The dataset’s computation consumed approximately 2 million CPU hours.

An experimental study was performed to find comparable or better adsorption energies than those found with DFT alone in OC20-Dense. The success rate metric, which measures the proportion of OC20-Dense systems where the ML+DFT adsorption energy is within 0.1 eV or lower than the DFT adsorption energy, was used to quantify this work. Success drops by about 5% when just using random configurations.

Success decreases much more noticeably when only considering heuristic setups. This result shows that random configurations can have a greater impact.

In this study, a new method called AdsorbML presents a series of accuracy-versus-efficiency trade-offs, with one well-balanced option finding the lowest energy configuration while achieving a 1387x increase in computing speed. The Open Catalyst Dense dataset, which has 87,045 different setups and about 1,000 different surfaces, is presented by the authors to standardize benchmarks. This study can be seen as a necessary first step in lowering the computational cost of DFT for computational chemistry in general, not just catalytic applications.

Look at the Paper and Project. All credit for this research goes to researchers on this project. Also, don’t forget to join our Reddit page and disagreement channelwhere we share the latest AI research news, cool AI projects, and more.