Design And Modeling of Oxide Catalysts by machine LEarning and atomistic Simulations

Información del proyecto

DAMOCLES

Identificador del acuerdo de subvención: 101108769

DOI

10.3030/101108769

Proyecto cerrado

Fecha de la firma de la CE 12 Abril 2023

Fecha de inicio 1 Septiembre 2023

Fecha de finalización 31 Agosto 2025

Financiado con arreglo a

• Marie Skłodowska-Curie Actions (MSCA)

Coste total

Sin datos

Aportación de la UE

€ 196 829,28

Coordinado por

AARHUS UNIVERSITET
Denmark

Periodic Reporting for period 2 - DAMOCLES (Design And Modeling of Oxide Catalysts by machine LEarning and atomistic Simulations)

Período documentado: 2024-09-01 hasta 2025-08-31

Resumen del contexto y de los objetivos generales del proyecto

A major fraction of chemical reactions needed by our society is carried out on heterogeneous catalysts, solid materials that enhance the rate of desired chemical reactions without being consumed during the process. Nowadays, heterogeneous catalysts are also expected to solve the challenges related to the diversification of energy sources and the reduction of pollutants and greenhouse gases. In this view, the catalytic process of CO2 hydrogenation can pave the way to sustainably produce chemicals and fuels from waste CO2 and green hydrogen, and mitigate global warming. To make processes such as CO2 hydrogenation economically advantageous, R&D in catalysis can rely on the discovery and optimization of catalytic materials based on the combination of experimental laboratory testing, spectroscopic analysis, and theoretical studies. Density functional theory (DFT) calculations, thanks to their accurate modeling of chemical bonds, can provide fundamental insight into the reaction mechanisms and guide laboratory experiments, avoiding the alternative procedure of trial and error. However, the modeling of heterogeneous catalysis from DFT is affected by the high computational demand of the calculations and the extreme complexity of the targeted system. In the last years, machine learning (ML) surrogate models have become standard tools to deal with the high computational cost of DFT calculations. Indeed, ML models trained on limited sets of DFT-calculated data can provide fast predictions on similar systems, drastically reducing the number of calculations required. Although large efforts have been devoted to applying predictive ML modeling for the design of catalytic materials, the attention has been concentrated mainly on metallic catalysts, while only a few works have addressed metal oxides. This represents a strong limitation to this research area, as metal oxides show great performances in several catalytic reactions important for our society, including O2 evolution reaction, selective catalytic reduction of NOx, and CO2 hydrogenation. Notably, indium and zirconium oxide catalysts showed superior catalytic performances in the CO2 hydrogenation to methanol, a reaction with great potential for the power-to-fuel production chain.
The overall objective of the DAMOCLES project is to advance computational catalysis by applying data-driven modeling (i.e. DFT combined with ML) to study and tailor metal oxide catalysts for CO2 hydrogenation, in search of new catalytic materials with improved performances (i.e. activity, and selectivity). By leveraging machine learning, the project seeks to model the underlying quantum mechanical principles of catalysis on oxide materials using small numbers of calculations, enabling the efficient and accurate prediction of other systems. This approach will help establish kinetic models of catalytic activity, guiding experimental efforts. In the outgoing phase of the project (first year), the postdoc Raffaele Cheula is hosted by Prof. John Kitchin at Carnegie Mellon University (USA). In the second year of the project, he returns to Aarhus University (Denmark) under the supervision of Prof. Mie Andersen.

Trabajo realizado desde el comienzo del proyecto hasta el final del período abarcado por el informe y los principales resultados hasta la fecha

In the outgoing phase of the DAMOCLES project, the postdoc Raffaele Cheula worked in the group of Prof. John Kitchin, at Carnegie Mellon University (CMU), Pittsburgh, PA. There, he learned how to use the machine learning potentials (MLPs) of the open catalyst project (OCP), a collaboration between CMU and Meta Fundamental AI Research. Then, he applied those MLPs to model reactions on the surfaces of metal and metal oxide materials. He focused on the study of transition state search techniques, required to calculate the kinetics of catalytic reactions. Along with this, he concluded a deep investigation of the CO2 hydrogenation to methanol on doped zirconium oxide, using atomistic simulations based on density functional theory (DFT).
In the return phase of the DAMOCLES project, the postdoc Raffaele Cheula worked in the group of Prof. Mie Andersen, at Aarhus University, Aarhus, Denmark. There, he focused on the application of DFT, descriptor-based graph machine learning models, and MLPs for the modeling of reactions on catalytic materials. Building on the methodologies developed during the first phase of the project, he designed automated workflows that integrate MLP-based simulations, transition-state search algorithms, and microkinetic modeling to study surface reactions. Then, he applied those workflows for the systematic investigation of large sets of materials, including doped zirconium oxides, perovskites, and single-atom alloys.
Additionally, during the two phases of the project, the postdoc Raffaele Cheula participated in studies on the identification of major sources of errors in the DFT datasets, the calculation of Gibbs free energies with MLPs, the extension of structure-dependent microkinetic modeling with inter-facet diffusion and reaction, and the use of MLPs and graph-based models to describe inverse catalysts.
During the two phases of the DAMOCLES project, the postdoc Raffaele Cheula participated in different workshops and conferences, and the results of the project were disseminated in scientific publications.

Avances que van más allá del estado de la técnica e impacto potencial esperado (incluida la repercusión socioeconómica y las implicaciones sociales más amplias del proyecto hasta la fecha)

An efficient and robust method for locating transition states in surface reactions was developed and optimized, offering a powerful tool for accelerating in-silico catalyst discovery. In addition, an automated workflow was created for exploring the kinetic mechanisms of surface reactions, enabling rapid prediction of the catalytic activity of both metal and oxide materials.
With these methodologies, a comprehensive study of CO2 hydrogenation on doped zirconium oxide was carried out, leading to a fundamental understanding of the kinetics of methanol formation and providing explanations for previously reported experimental observations. Then, a systematic catalyst design study was performed by screening all elements of the periodic table as dopants, identifying the most promising candidates for improving catalytic activity and selectivity. Additionally, a detailed dataset of adsorption and transition-state energies on doped zirconium oxide was generated. The methodologies developed were also applied to the study of perovskite materials, which hold great potential as tunable oxide catalysts for CO2 conversion reactions.
Overall, the results of the project significantly advance the field of computational catalysis, particularly in the understanding, design, and modeling of metal oxide and perovskite catalysts.

picture.png