Simulations of reactions on catalyst surfaces have augmented the rationale for the design of catalytic materials at the molecular level. For in-silico high throughput screening, this design principle is often simplified by the identification of a molecular level reactivity descriptor such as the location of the d-band center from the fermi level of the transition metal catalysts. While ab-initio simulations have demonstrated success in predicting reactivity trends on an array of transition metal, bimetallic, and oxide catalysts, the computationally intensive requirement for building a large dataset of reaction energetics has often limited its applicability. Herein ab-initio simulations are augmented with a machine learning (ML) approach, to speed up this in-silico screening process. The ML method applied here is the Gradient Boosting Regression (GBR) algorithm, which is showing applicability in working with small datasets, since massive data generation from ab initio simulations is an arduous task. The applicability of the GBR algorithm for in-silico catalyst screening is demonstrated in two cases. In the first case, the GBR algorithm is applied to predict binding energies of carbon and oxygen on transition metal surfaces, to build microkinetic models (MKMs) for estimating the rates of ethanol decomposition and dehydrogenation reactions. The ML-enabled MKM is further utilized to design catalytically active bimetallic and single atom alloys. In the second case, the same algorithm is implemented for screening heteropolyacid (HnM12XO40) catalysts, while changing their addenda M atom from the d-block elements and the hetero X atom from the p-block elements, so as to predict their deprotonation energies, which are often applied as reactivity descriptors for acid-catalyzed reactions. In both cases, periodic table properties of the elements are taken as input features for target descriptors—ensuring good predictions on a small dataset. This is enabling a more comprehensive description of catalyst reactivity. Moreover, the algorithm is observed to efficiently finds the error while developing the model. It understands the error in the initial model and builds up further models based on the errors. This process finally yields a model ensemble, providing an improvised fit to the dataset.

Interpreting catalyst reactivity descriptors using a machine learning approach with small datasets

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.