Discovering Catalytic Reaction Networks Using Deep Reinforcement Learning from First-Principles

Determining the reaction pathways, which is central to illustrating the working mechanisms of a catalyst, is severely hindered by the high complexity of the reaction and the extreme scarcity of the data. Here, we develop a novel artificial intelligence framework integrating deep reinforcement learning (DRL) techniques with density functional theory simulations to automate the quantitative search and evaluation on the complex catalytic reaction networks from zero knowledge. Our framework quantitatively transforms the first-principles-derived free energy landscape of the chemical reactions to a DRL environment and the corresponding actions. By interacting with this dynamic environment, our model evolves by itself from scratch to a complete reaction path. We demonstrate this framework using the Haber-Bosch process on the most active Fe(111) surface. The new path found by our framework has a lower overall free energy barrier than the previous study based on domain knowledge, demonstrating its outstanding capability in discovering complicated reaction paths. Looking forward, we anticipate that this framework will open the door to exploring the fundamental reaction mechanisms of many catalytic reactions.

阐明催化剂的工作机制，核心在于确定其反应路径，但反应体系的高度复杂性与相关数据的极度匮乏，严重阻碍了这一目标的达成。本研究构建了一种全新的人工智能框架，将深度强化学习（deep reinforcement learning, DRL）技术与密度泛函理论（density functional theory, DFT）模拟相结合，实现了从零起步对复杂催化反应网络的自动化定量搜索与评估。本框架将基于第一性原理计算得到的化学反应自由能面，定量转化为深度强化学习环境与对应动作空间。通过与该动态环境交互，模型可从零起步自主演化，最终得到完整的反应路径。本研究以最具催化活性的Fe(111)晶面上的哈伯-博施法（Haber-Bosch process）为实例，验证了该框架的有效性。本框架发现的新路径，其总自由能垒低于此前基于领域知识得到的研究结果，充分证明了该框架在复杂反应路径发掘方面的卓越能力。展望未来，我们期望该框架能够为众多催化反应的基础反应机制探索打开全新的研究大门。