Machine Learning Predictions of Factors Affecting the Activity of Heterogeneous Metal Catalysts

Abstract

The ultimate goal in heterogeneous catalytic science is to accurately predict trends in catalytic activity based on the electronic and geometric structures of active metal surfaces. Such predictions would allow the rational design of materials having specific catalytic functions without extensive trial-and-error experiments. The d-band center values of metals are well known to be an important parameter affecting the catalytic activity of these materials, and activity trends in metal surface catalyzed reactions can be explained based on the linear Brønsted-Evans-Polanyi relationship and the Hammer-Nørskov d-band model. The present work demonstrates the possibility of employing state-of-the-art machine learning methods to predict the d-band centers of metals and bimetals while using negligible CPU time compared to the more common first-principles approach.

Statistical Analysis and Discovery of Heterogeneous Catalysts Based on Machine Learning from Diverse Published Data

“…Two of these methods involve linear regression methods (Lasso, Ridge), and the other five methods involve nonlinear regression methods [more specifically, two kernel methods (KRR, SVR) and three tree ensemble methods (RFR, GBR, ETR)]. This set of ML models covers a wide spectrum of model types, which can reveal the relevant aspect of diverse data as illustrated in previous research on data‐driven predictions for DFT‐calculated values such as d‐band centers and adsorption energies . Widely used implementations of scikit‐learn (version 0.19.1) were employed for all ML models except XGB, and for XGB the original implementation (version 0.81) of XGBoost was used.…”

“…Several successful examples have already been reported for organic chemistry reactions, including those that involve homogeneous catalysts . However, the applicability of ML predictions for heterogeneous catalysis have been limited mainly to computationally determined values such as band gaps, d‐band centers, and adsorption energies . For the practical use of ML for discovering new solid catalytic materials, not only first‐principles calculated values but also experimental values for specific catalytic reactions are needed, especially in heterogeneous catalysis because an adequate theoretical model for heterogeneous catalysis is not available.…”

The effect of SnO2(110) supports on the geometrical and electronic properties of platinum nanoparticles

“…2a [66][67][68]. It was proved that the relationships between adsorption energy and d-band center do not account for the effect of less coordinated atoms, such as the ones located at the vertices and edges of nanoparticles, on the adsorption, especially for small cluster particles that do not expose well-defined planes [72][73][74]. Considering the experimental value of − 2.86 eV of Pt 3 Co surfaces, the calculated values of the d-band centers of the outer-shell atoms of the Pt-nanoparticles considered in this study point out to a stronger adsorption of oxygenated species; O 2− , OH − , H 2 O, H 2 O 2 , etc.…”

Machine Learning for Computational Heterogeneous Catalysis

“…As much of the transition metal chemistry is defined by the alignment of the metal's d-band center and the adsorbate frontier orbitals, it is indeed a valuable feature. [50,53,54] Other features that can be derived from the electronic structure are, for example, band gaps, electron mobilities and number of states at the Fermi level. Recently, the average 2p-state energy of surface oxygen atoms was shown to be a strong descriptor for the oxygen reactivity at metal and metal-oxide surfaces.…”

“…[101] Takigawa et al found that Gaussian process regression showed superior accuracy compared to five linear models and five nonlinear models with a prediction error of the d-band center below 0.2 eV compared to the density functional theory benchmark. [54] Gaussian processes are very powerful, and hyperparameters can be neatly optimized by maximizing the marginal likelihood. Furthermore, Gaussian processes predictions come with estimated errors of that prediction, which are empirical confidence intervals.…”