Towards Transferable Machine Learning Interatomic Potentials for Reactive Organic Chemistry in Solution

液相反应是合成有机化学研究和工业的核心。溶剂以复杂的方式影响反应机理和速率，包括非共价相互作用和质子转移。溶剂效应尚未被完全理解。计算机模拟可以帮助理解溶剂效应，通过自由能计算和处理整个系统在高水平的理论。这通常是使用从头算分子动力学（AIMD）来完成的，但由于计算成本高，应用范围有限。机器学习原子间势（MLIPs）正在成为精确凝聚态模拟的常用工具，与AIMD相比，它延长了模拟的时间和长度尺度，而不会显著降低精度。开发MLIP的关键挑战是对相关PES进行有效采样。已经有人尝试建立一般的ML电位，也有人尝试为特定的凝聚相反应创建mlip。mlip的高灵活性为FFs打开了大门，可以准确地代表由选定的化学元素所跨越的大部分化学成分。目前还没有尝试建立一个通用的反应性MLIP来研究溶液中的有机反应。我们将开发一种策略来有效地采样代表所选溶液相反应系统PES的配置。然后，我们将推广这种方法来构建一个通用的MLIP，该MLIP可以在不同的反应物-溶剂系统中实现合理的准确性，并且不需要为每个新应用重新训练。最后，我们希望利用我们的模型进行与现代有机化学研究相关的反应体系的计算。

Reactions in liquid phase are central to research and industry involving synthetic organic chemistry. Solvents influence reaction mechanism and rate in complicated ways involving non-covalent interactions and proton transfers. Solvent effects are not yet completely understood. Computer simulation can aid the understanding of solvent effects through free energy calculations and treatment of the full system at a high level of theory. This is commonly done using ab initio molecular dynamics (AIMD), which has a limited scope of application due to high computational cost. Machine learning interatomic potentials (MLIPs) are becoming a common tool for accurate condensed phase simulation, extending the time- and length-scale accessible to simulation without significant loss of accuracy compared with AIMD. The key challenge to developing an MLIP is the efficient sampling of relevant PES. There have been attempts at building a general ML potential as well as attempts at creating MLIPs for a specific condensed phase reaction. The high flexibility of MLIPs opens the door to FFs that accurately represent a large fraction of the chemistry spanned by a selection of chemical elements. There has not been attempt to build a general reactive MLIP for studying organic reactions in solution. We will develop a strategy to efficiently sample configurations that represent the PES of a chosen solution-phase reactive system. We will then generalise this approach to build a general-purpose MLIP that achieves reasonable accuracy across different reactant-solvent systems and does not need to be re-trained for each new application. Finally, we hope to use our models to carry out calculations of reactive systems relevant to modern organic chemistry research.