Building better molecular models through artificial intelligence

绝大多数对液体和溶液的模拟研究使用的是不可极化的成对势能模型，或力场，这种模型可以追溯到20世纪90年代。它们的相对简化的性质使得在化学、生物、工程和材料科学等领域取得了巨大的进步，使得可以在分子水平上模拟高度复杂的系统。然而，尽管它们取得了成功，但越来越明显的迹象表明，它们已经超过了到期日期，这表明在不可极化的分子模型中存在一些本质上缺失的物理学。例如，对极化效应的忽视一直被认为是经典模型中的“房间里的大象”，而更复杂、更耗时的极化模型尚未在预测能力方面产生所需的改进，特别是对混合物和溶液。更广泛地说，成对力场主要基于“原子类型”的概念，即分子是通过基于化学直觉将一组相互作用的单元组装在一起来构建的。然而，已有研究表明，使用这样严格的“分子”组成定义可能会导致性质预测的不一致。最后，我们仍然不清楚什么是适合足够精确的力场的最佳实验数据集。这个项目的主要目标是彻底改变这个问题--也就是说，我们的目标不是受化学直觉、有限的实验数据和传统方法的限制，而是使用人工智能(AI)方法来“教”我们如何更好地建立分子模型。该项目将解决一些基本的科学问题，例如：定义一个“分子”的最佳方式是什么，即定义它与周围环境相互作用的一组数学函数？分子的化学连通性在多大程度上对其相互作用起作用？在广泛的热力学条件下建立具有强大预测性能的力场所需的最小实验数据集是什么？在力场参数化中，纯液体数据和混合数据之间的正确平衡是什么？如何在不使用昂贵的可极化模型的情况下将极化效应隐含地包括在模型参数中？通过这样做，我们将为下一代分子模型奠定基础，这些模型可以可靠地用于多种技术应用的分子和材料的预测设计，包括：用于能量存储的电解质系统的设计；用于碳捕获的多孔材料的设计；药物的设计和制造。

The vast majority of simulation studies of liquids and solutions make use of non-polarisable pair-wise potential models, or force fields, that date back to the 1990s. Their relatively simplified nature has enabled immense progress to be made in fields like chemistry, biology, engineering and material science, allowing for highly complex systems to be simulated at the molecular level. Despite their successes, however, there are increasingly clear signs that they have exceeded their expiry date, suggesting that there is some essential missing physics in non-polarisable molecular models. For example, the neglect of polarisation effects has long been recognised as the "elephant in the room" of classical models, while more complex and time-consuming polarisable models have not yet yielded the required improvements in prediction ability, particularly for mixtures and solutions. More generally, pair-wise force fields are mostly based on the concept of "atom types", whereby molecules are built by assembling together a set of interacting units based on chemical intuition. However, it has been shown that using such a rigid definition of what constitutes a "molecule" can lead to inconsistencies in property predictions. Finally, it is still unclear what is the best set of experimental data required to fit a sufficiently accurate force field.The main objective of this project is to turn this problem on its head - i.e. instead of being constrained by chemical intuition, limited experimental data and conventional approaches, we aim to use artificial intelligence (AI) methods to "teach" us how to better build molecular models. The project will address fundamental scientific questions such as: What is the best way to define a "molecule" as a set of mathematical functions that define its interactions with the surrounding environment? To which extent does the chemical connectivity of the molecule play a role in its interactions? What is the minimal set of experimental data that is required to build a force field with strong prediction performance over a wide range of thermodynamic conditions? What is the right balance between pure-liquid and mixture data in force field parametrisation? How can polarisation effects be implicitly included in the model parameters without using expensive polarisable models? In so doing, we will lay the foundation for the next generation of molecular models that can be reliably used in predictive design of molecules and materials for multiple technological applications, including: design of electrolyte systems for energy storage; design of porous materials for carbon capture; drug design and manufacture.