Machine learning in catalysis
John R. Kitchin, “Machine learning in catalysis”, Nature Catalysis 1(4), 230–232 (2018). https://doi.org/10.1038/s41929-018-0056-y
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
This is a three-page Comment rather than a research article, so it reports no new data. The summary below describes its arguments and the published work it uses as examples.
Question
The Comment asks how machine learning (ML) was being used in catalysis research as of early 2018, and what opportunities and challenges lie ahead (Kitchin 2018). It contrasts classical model building, in which physical principles such as conservation laws, thermodynamics or quantum chemistry give a formula whose parameters are fit to data, with ML, in which flexible and often nonlinear models are trained on reference data to output the desired quantity. The author notes that ML was applied to experimental catalysis data decades earlier, and attributes its current renaissance to open-source libraries (scikit-learn, TensorFlow, PyTorch, Chainer), cheap GPU hardware and cloud computing, which together lowered the barrier for non-experts.
Approach
The piece is a survey organized around two main roles for ML in catalysis, machine-learned atomistic potentials and the prediction of catalytic properties, followed by an outlook. Each role is illustrated with a handful of recent studies from other groups and the author’s own.
Key findings
On atomistic potentials, the Comment argues that density functional theory (DFT) is too expensive when thousands to hundreds of thousands of calculations are needed (screening, free energies, Monte Carlo or molecular dynamics), while classical potentials are fast but may be inaccurate and hard to improve systematically. Potentials learned from DFT fill that gap, and three open-source codes for building them are named: Amp, aenet and Prophet, alongside Gaussian process and kernel ridge regression approaches common in chemistry and materials. Examples include a neural network potential that allowed molecular dynamics with DFT accuracy of the full dissociation of N2 on Ru(0001), including coupling of surface phonons to adsorbate vibrations, with results agreeing with experiment; neural network potentials used to simulate infrared spectra; and a potential for alloy nanoparticles in water that covered particles of 147 to 1,415 atoms, giving composition profiles, surface energies, temperature-dependent segregation for 1.0 and 1.5 nm particles, and the structure of interfacial water (Kitchin 2018).
On property prediction, the Comment contrasts an ML model of the d-band center that used up to nine elemental descriptors (atom size, electronegativity, density and others), with moderately accurate results from as few as six, against the author’s earlier physics-based model whose only input was active-site geometry plus fitted d-orbital sizes. The ML model predicts well but does not explain why its inputs matter. By contrast, work by Xin and coworkers used physically motivated descriptors to screen CO adsorption on core–shell alloys and in doing so identified contributions from d-band shape and sp-band filling (notably for coinage metals) that the simple d-band model omits. Further examples cover estimating nanoparticle site reactivity through local similarity to DFT-computed surface sites or through decision trees and neural networks on facet and structural features; uncertainty-guided DFT exploration of syngas chemistry on Rh(111), where the network has more than 2,000 possible pathways; neural networks that predict the sensitivity of transition metal complex spin-state splitting to the exchange-correlation functional; and ML that solves the inverse problem of inferring structure from measured XANES spectra by training on thousands of simulated structure and spectrum pairs (Kitchin 2018).
In the outlook, the author points to data-rich experiments, text mining of the catalysis literature (citing materials science efforts that mined 640,000 articles on oxide synthesis and about 200 papers on metal organic frameworks), image analysis of microscopy data, generative models that encode molecules in a continuous space, optimize there and decode back (with no solid-state or nanoparticle analogue known to the author), and ML acceleration of Kohn–Sham DFT. The challenge singled out as most significant is reporting: models with hundreds of parameters, many hyperparameters, data processing pipelines and training sets of hundreds to tens of thousands of points exist as programs rather than equations, and narrative Methods sections and current supporting information formats cannot capture them well enough for reproduction (Kitchin 2018).
Significance
The Comment is an early programmatic statement from the group on ML in catalysis. It stresses that feature choice needs domain knowledge, that ML can reveal physics as well as accelerate screening, and that the community needs new data and code sharing practices and education in ML for scientists. Several of its themes (learned potentials, interpretable property models, inverse problems and reproducible reporting) recur in later work from the group.