Automating reaction condition prediction for supramolecular organic assemblies

超分子合成是组装复杂分子的强大策略，但通过靶向设计来实现这一点具有挑战性。虽然使用化学直觉在靶向特定分子方面取得了一些成功，但其他合成变量（如溶剂，温度和浓度）如何控制反应结果的知识是有限的，并且可以说是偶然的。我们希望采用数据驱动的方法来预测反应结果，但由于缺乏足够高数量的高质量数据而受到阻碍。在这里，我们建议使用高通量自动化进行大量的反应筛选，使用市售和简单的合成可行的前体。这将填补数据中的差距，在不同的前体组合中产生数千个数据点，用于动态共价化学、溶剂选择、温度选择和反应浓度，所有这些都将记录在电子实验室笔记本中。根据现有数据，我们将使用监督机器学习方法，尝试首先预测成功的“清洁”反应的最佳合成条件，然后预测反应结果（即分子量）。我们还将在可能的情况下使用Jelfs组内可用的软件应用原子模型，以确定哪些命中最有希望扩大规模，并调查观察到的设计规则背后的物理原因。

Supramolecular synthesis is a powerful strategy for assembling complex molecules, but doing this by targeted design is challenging. Whilst there is some success in targeting specific molecules using chemical intuition, knowledge of how other synthetic variables such as solvent, temperature, and concentration, control the reaction outcome is limited, and arguably left to chance. We would like to move towards a data-driven approach to predicting the reaction outcomes but are hampered by the lack of high quality data in a sufficiently high quantity. Here, we propose to use high-throughput automation to carry out large numbers of reaction screens using both commercially available and simple synthetically viable precursors. This will fill the gap in data by generating thousands of data points across different precursor combinations for dynamic covalent chemistry, solvent choice, temperature choice, and reaction concentration, all of which will be recorded in an electronic lab notebook. With the available data, we will use supervised machine learning approaches to attempt to first predict the best synthetic conditions for a successful, "clean" reaction, and then to predict the reaction outcome (i.e. molecular mass). We will also apply atomistic modelling where possible using available software within the Jelfs group, to both determine which hits are the most promising to scale up and to investigate the physical reasons behind observed design rules.