Closed-loop discovery of materials and catalysts with automation and machine learning

在化学研究中使用新技术、机器人和自动化的动力越来越大。自动化平台可以在很短的时间内进行数千项实验。还可以使用配备高通量自动进样器的套件快速收集这些反应产物的表征数据，包括质谱学和核磁共振光谱等关键分析技术。然而，发现过程中的瓶颈随后变成了对这些海量数据的分析--这些数据的解释和分配在很大程度上仍然是由人来执行的，而核磁共振谱的分析非常耗时。这可能会导致平台长时间闲置，而人类正在分析结果并确定哪些产品是制造的。该项目将专注于开发软件，自动分析在高通量筛选工作流程中收集的核磁共振数据。这将包括处理原始数据，挑选峰，确定多重态，并将这些与计算化学模拟计算的化学位移进行比较，最终允许确认混合物中存在哪些产品或进行结构指定。将这些自动化平台与开发的软件(包括人工智能算法)结合起来，有望实现“闭环”发现，即由软件分析一组初始反应，然后根据结果，算法决定下一组要尝试的实验，然后在平台上运行，例如优化材料的性质或催化剂的选择性。这将涉及试验各种优化方法，例如使用进化算法或贝叶斯优化。开发的方法将具有广泛的适用性，并将针对各种问题进行测试，包括筛选各种催化剂的选择性，以及鉴定成功合成有前景的分子材料。该项目符合沙坑中的数字化/新工具/快速分析主题。

There is an increasing drive to use new technologies, robotics, and automation in chemical research. Automated platforms can allow thousands of experiments to be carried out within a very short time. Characterisation data of the products of these reactions can also be rapidly collected using kit equipped with high-throughput autosamplers, including key analytical techniques such as mass spectrometry and NMR spectroscopy. However, the bottleneck in the discovery process then becomes the analysis of this huge amount of data - interpretation and assignment of this data is still largely carried out by a human, and the analysis of NMR spectra is very time consuming. This can lead to platforms lying idle for large periods of time while a human is analysing the outcome and determining which products were made. This project will focus on the development of software that automates the analysis of the NMR data collected during high-throughput screening workflows. This will include processing the raw data, peak picking, determining multiplets, and comparing these to chemical shifts calculated by computational chemistry simulations, finally allowing either confirmation of what products are present in a mixture or structural assignment. Combining these automated platforms with the developed software, including artificial intelligence algorithms, holds the promise of "closed loop" discovery where an initial set of reactions are analysed by software, and then based on the outcome, the algorithm decides on the next set of experiments to try and then run on the platform, for example to optimise the properties of a material or the selectivity of a catalyst. This will involve trialing a variety of optimisation approaches, such as using evolutionary algorithms or Bayesian optimisation. The approach developed will have wide applicability and will be tested against a variety of problems, including screening for the selectivity of a wide range of catalysts and for the identification of the successful synthesis of promising molecular materials. The project fits in the Digitisation/New Tools/Rapid Analysis theme from the Sandpits.