Deep-learning-enhanced simulation of plasmonic CO2 catalysis

Photosynthesis is an important reaction in nature, which is used by plants to convert the greenhouse gas carbon dioxide (CO2) into sugar molecules. This reaction is driven by sunlight and requires no additional energy source, making it very attractive for scientists to create an artificial counterpart to this reaction. An exciting new field of research that has the potential to achieve such an environmentally friendly conversion of CO2 into higher-valued resources is photoplasmonic catalysis. This field of research takes advantage of the ability of noble metals such as gold or silver in the form of nanoparticles to harness the energy when exposed to light. Recently, experimental studies have shown that the molecule CO2 can be destroyed by such plasmonic metal nanoparticles when exposed to light. Unfortunately, the mechanisms underlying this reaction are not yet clear and therefore it is not known what properties a perfect plasmonic metal nanoparticle must have in order to efficiently convert CO2. The idea behind this project is to develop a new method based on machine learning to study the light-induced reactions of molecules on metal surfaces and to compute the reaction rates of catalytic conversion. Machine learning is part of the broader concept of artificial intelligence, which means that such methods can be used to learn from data and can help to find relationships within data with computational efficiency and the accuracy of the underlying data. This concept is particularly interesting for photoplasmonic catalysis: First principle methods are usually too expensive to compute observables at the scales needed to complement experiments. However, such methods can be used to generate reference data for machine learning models, which can be subsequently used to simulate experimentally measurable quantities, such as reaction rates. In this way, computationally efficient simulations can be enabled, which can help to develop a better understanding of photocatalytic conversion. In order to make this project feasible, it strongly interlaces the knowledge of three research groups: The Computational Surface Chemistry group of Assoc. Prof. Dr. Reinhard Maurer at the University of Warwick, the Photochemistry research group of Dr. Philipp Marquetand at the University of Vienna and the Computer Science and Machine Learning research group at the Technical University of Berlin with Univ. Prof. Dr. Klaus-Robert Müller and Dr. Michael Gastegger. A close interdisciplinary collaboration with experimental research groups is aimed for in different phases of this project. By combining the expertise of the different research groups, the necessary conditions for a new method for the investigation of photoplasmonic catalysis shall be created. The photocatalytic conversion of CO2 on metal surfaces will serve as a test reaction for this method and may contribute to the development of novel catalysts to efficiently convert CO2 into higher-value resources.

Light-matter interactions play a role in everyday life, whether it is seeing with the human eye or photosynthesis. The latter reaction in particular has the potential to provide a solution to one of society's most pressing challenges, which is global warming, as it describes the ability of plants to convert the greenhouse gas carbon dioxide into sugar molecules using only light as an energy source. Photoplasmonic catalysis has emerged as a promising research field to mimic photosynthesis and enable environmentally friendly conversion of carbon dioxide into value-added resources using noble metal nanoparticles such as gold or silver that can harness light and convert it into chemical energy at their interfaces. However, in order to develop an artificial counterpart to photosynthesis, it is important to understand the underlying mechanisms of photoplasmonic catalysis. Currently, the major limitation hindering a thorough study of light-matter interactions of molecules, materials, and hybrid interfaces is the high complexity and computational cost associated with quantum chemical calculations. The project "Deep-learning-enhanced simulation of plasmonic CO2 catalysis" aims to provide theoretical tools based on artificial intelligence to better describe light-matter interactions of molecules (at interfaces) and to efficiently design new systems with specific properties. Therefore, we developed a physics-inspired deep neural network that can describe the interactions of molecules with light. As a proof of principle, we focused on functional organic molecules suitable for organic electronics, such as those used in photovoltaics. This method is applicable for screening millions of molecules. However, to design the next generation of optoelectronics, a targeted design is required. Thus, we developed a second, generative, machine learning method that can learn structural arrangements of molecules and design novel molecules. By screening newly predicted molecules with the previously developed deep neural network model, we were able to bias the generative model to produce only the most relevant molecules. In this way, molecules with properties far outside the original property space could be generated. Because of the large space that needs to be covered when extending the approach to molecules on nanoparticles, we developed a method to explicitly account for long-range effects. In this way, molecules on surfaces could be screened and adsorption processes could be monitored. The methods developed will provide the scientific community with tools to describe extended systems such as nanoparticles or hybrid interfaces computationally efficiently and with near experimental accuracy. In combination with the generative model and the targeted design method, economic and technological applications can be achieved, at least in principle, for instance by developing new drugs or materials.