Deep learning enabled simulation of plasmonic photocatalysis

等离子体光催化为更可持续和高效的化学转化提供了一条有前途的途径。金属催化剂可以通过激发电子来利用光，有选择地将能量传递给分子并促进化学反应。结果是增加了反应选择性和减少了不需要的副产物。这种非常规的化学形式涉及光、电子激发和分子运动的复杂耦合，其细节仍在激烈的争论中。等离子体光催化的理论研究预测反应概率作为催化剂组成、形状和光照的函数，受到实际系统从头算分子动力学模拟的计算成本的限制。该项目旨在开发和应用新的分子模拟方法，这些方法既准确又可扩展，足以研究金属催化剂上的光驱动化学反应。该项目的主要飞跃是开发基于消息传递神经网络的深度机器学习（ML）电子结构替代模型，该模型可以以从头计算计算成本的一小部分提供预测。实现这一点将使计算成本与预测精度解耦。这些ML替代模型将与非绝热分子模拟方法和介观光物质相互作用模型相结合，通过在各种反应条件下平均数千个反应事件来模拟实验可测量的反应概率。我们将通过模拟等离子体光增强氢的演化，以及一氧化碳和二氧化碳的减少作为关键设计参数的函数来展示我们的方法的转换能力。这个项目将超越目前的技术水平，改变我们设计等离子体催化剂材料的能力，仔细审查机制建议，并指导关键催化反应的实验。

Plasmonic photocatalysis offers a promising route to more sustainable and efficient chemical transformations. Metal catalysts can harness light via excitation of electrons which selectively transfer energy to molecules and promote chemical reactions. The result is an increase of reaction selectivity and a decrease of unwanted side products. This unconventional form of chemistry involves intricate coupling of light, electronic excitations, and molecular motion, the details of which are still under intense debate. The theoretical study of plasmonic photocatalysis to predict reaction probabilities as a function of catalyst composition, shape, and light exposure is limited by the computational cost of ab initio molecular dynamics simulations of realistic systems. This project seeks to develop and apply new molecular simulation methods that are both accurate and scalable enough to study light-driven chemical reactions on metal catalysts. The major leap this project will take is to develop deep machine learning (ML) surrogate models of electronic structure, based on message-passing neural networks that provide predictions at a fraction of the computational cost of ab initio calculations. Achieving this will decouple computational cost from prediction accuracy. These ML surrogate models will be combined with nonadiabatic molecular simulation methods and mesoscopic light-matter interaction models to enable the simulation of experimentally measurable reaction probabilities by averaging over thousands of reaction events at various reaction conditions. We will showcase the transformational capabilities of our methodology by simulating plasmonic light-enhancement of hydrogen evolution, and carbon monoxide and carbon dioxide reduction as a function of key design parameters. This project will go beyond the state of the art by transforming our ability to design plasmonic catalyst materials, to scrutinize mechanistic proposals, and to guide experiments for key catalytic reactions.