Exploring Chemical Compound Space with Machine Learning

准确预测化合物空间的分子性质是化学和制药工业中合理设计化合物的关键因素。因此，主要的挑战之一是如何以适中的计算成本（每个分子毫秒级或更快）实现CCS中分子特性的定量计算。然而，目前只有高水平的量子化学计算才能产生预期的“化学精度”（1千卡/摩尔），这是预测\textit{硅}理性分子设计所需的，每个分子可能需要几天的时间。机器学习（ML）方法已被成功地用于将求解复杂物理微分方程的问题映射到统计模型。在这个项目中，我们将评估有效的ML方法在应用于预测量子化学计算获得的不同分子性质时的能力。主要的焦点将是预测分子能量，然而，同样的想法可以在稍后的阶段用于预测激发态性质，如极化率，电离势或电子亲和。我们的最终目标是使分子能量的预测接近“化学精度”，而成本只是电子结构计算的一小部分。实现这一目标将使我们能够合理地探索和分析CCS的结构和维度。该项目的预期结果是：(a)使用最优ML模型对CCS进行物理分析（探索），以期识别重要的分子类别并了解CCS的维度。(b)对使用ML技术预测分子性质的可行性（能力和局限性）进行严格评估，最后(c)用不同理论水平计算的各种分子的分子性质和激发态性质的数据集。

The accurate prediction of molecular properties in the chemical compound space (CCS) is a crucial ingredient toward rational compound design in chemical and pharmaceutical industries. Therefore, one of the major challenges is to be enable quantitative calculations of molecular properties in CCS at moderate computational cost (milliseconds per molecule or faster). However, currently only high level quantum-chemical calculations, which can take up to several days per molecule, yield the desired 'chemical accuracy' (1~kcal/mol) required for predictive \textit{in silico} rational molecular design.Machine learning (ML) methods have been successfully used to map the problem of solving complex physical differential equations to statistical models. In this project, we will assess the capability of efficient ML methods when applied to the prediction of different molecular properties obtained with quantum chemistry calculations. The main focus will be on predicting molecular energies, however the same ideas can be employed at a later stage to predict excited state properties, such as polarizability, ionization potential or electron affinity.Our final aim is to enable predictions of molecular energies close to 'chemical accuracy' at a small fraction of cost of electronic structure calculations. Achieving this goal will allow us to rationally explore and analyze the structure and dimensionality of CCS.The expected results of this project are: (a) a physical analysis (exploration) of CCS using optimal ML models, with an outlook to identify important classes of molecules and understand the dimensionality of CCS. (b) A rigorous assessment of the feasibility (capabilities as well as limitations) of using ML techniques for the prediction of molecular properties, and finally (c) a dataset of molecular properties and excited state properties for a wide variety of molecules computed with different levels of theory.