Open Challenges in Developing Generalizable Large-Scale Machine-Learning Models for Catalyst Discovery

Abstract

The development of machine-learned potentials for catalyst discovery has predominantly been focused on very specific chemistries and material compositions. While they are effective in interpolating between available materials, these approaches struggle to generalize across chemical space. The recent curation of large-scale catalyst data sets has offered the opportunity to build a universal machine-learning potential, spanning chemical and composition space. If accomplished, said potential could accelerate the catalyst discovery process across a variety of applications (CO2 reduction, NH3 production, etc.) without the additional specialized training efforts that are currently required. The release of the Open Catalyst 2020 Data set (OC20) has begun just that, pushing the heterogeneous catalysis and machine-learning communities toward building more accurate and robust models. In this Perspective, we discuss some of the challenges and findings of recent developments on OC20. We examine the performance of current models across different materials and adsorbates to identify notably underperforming subsets. We then discuss some of the modeling efforts surrounding energy conservation, approaches to finding and evaluating the local minima, and augmentation of off-equilibrium data. To complement the community’s ongoing developments, we end with an outlook to some of the important challenges that have yet to be thoroughly explored for large-scale catalyst discovery.

Towards predictive design of electrolyte solutions by accelerating ab initio simulation with neural networks

“…The underlying theoretical details of these methods are beyond the scope of this review and we refer the reader to several excellent reviews of these methods that have recently been published for more detail. 45–51 At a high level these methods work by defining a mapping between atomic coordinates to energies and forces (occasionally virial tensors also). This mapping contains a large number of parameters (weights and biases) that can be systematically adjusted to minimise the error on a set of training data, combined with an algorithm to systematically optimise the parameters (Backpropagation).…”

The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts

“…Pursuant to OC20, non-DFT distance based metrics like ADwT struggle to correlate well with the practical DFT metrics. 10 Both FbT and AFbT results indicate the models need significant improvement to achieve the level of accuracy needed for practical applications. Does OC22 benefit OC20?…”

“…[47][48][49][50][51][52] Models developed for OC20 have shown great progress on their proposed tasks. 10,[15][16][17]19,20 In all of OC20's tasks, energies were referenced to represent adsorption energy. While advantageous for screening purposes, this referencing, however, implicitly limited models to only studying adsorbate+slab combinations and not any one in isolation.…”

“…The dataset of slabs in OC22 is constructed from a set of 51 elements shown in Figure 10 = 1275 possible unary and binary oxides respectively. We considered all transitions metals up to the 5d group with the exception of Tc due to its radioactivity (29 metals), all alkali and alkaline earth metals up to Fr (10), the lanthanides of Ce and Lu (2), and the p-block metals and metalloids: except Te and up to Po (11). We queried the materials project for materials with the top five lowest energy above hull and less than 150 atoms for all unary and binary oxides composed of these elements.…”

“…to the complexities above. The release of the OC20 dataset helped enable rapid advances in the accuracy and generalizability of Graph Neural Network (GNN) models, 10 with decreases of 55+% in the key S2EF metrics in the first two years. Initial baseline models like CGCNN 11 and SchNet 12 focused on local environment representations.…”