Library Synthesis in Flow Made Easier: DoE & ML

通过高通量实验(HTE)和自动筛选级联快速优化反应条件以应对可能的合成挑战的能力长期以来一直是行业标准工作流程。然而，在对单一化合物的反应进行优化的过程中做出的许多决定(已实现的“局部”合成最大值)在应用于研究项目环境(“全局”或多个合成最大值)时，如果应用于给定合成序列所需的化学多样性的广度，则会产生不可预见的后果。因此，合成序列需要从一开始就重新设计，这可能会导致低效率的过程(时间、成本、环境负担等)。我们的加速方法旨在将用于增强实验优化的自我优化自动化平台(例如SNOBFIT左上角)与先正达的化学案例研究结合起来。我们的目标是解决流动中文库合成的局部与全局最大值问题。我们通过在主成分分析中使用选定的分子描述符来确定反应速度、化学产率和选择性的敏感性，然后通过实验设计(DoE)方法来实现这一点。所得到的统计模型将用于生成合成条件的初始最佳猜测起点。该模型将定期更新，添加通过自动优化产生的试剂和反应条件的额外筛选数据。统计模型和实验结果之间的这种反馈循环将随着每次迭代而提高精度，从而产生一个预测工具，用于绘制与初始模型外部的起始材料、催化剂和溶剂的反应性。这种结构-性质/反应性关系与基于统计响应面(DOE)的方法相结合，然后将与该过程的数字孪生兄弟相结合，从而实现对反应类别进行整体优化的方法，包括开发决策支持框架。

The ability to optimise reactions conditions rapidly for possible synthetic challenge through high throughput experimentation (HTE) and automated screening cascades have long been industry standard workflows. Yet, many decisions made during the optimisation of a reaction towards a single compound ('local' synthetic yield maxima achieved) have unforeseen consequences when applied to the breadth of chemical diversity required from a given synthetic sequence when applied in a research project setting ('global' or multiple synthetic maxima required). Thus, synthetic sequences need to be re-designed from the beginning, which can lead to inefficient processes (time, cost, environmental burden etc.). Our accelerated approach aims to couple the self-optimising automated platforms for augmented experimental optimisation (e.g. SNOBFIT top left) with chemical case studies from Syngenta. Our aim is to address this 'local vs global' maxima problem for library synthesis in flow.We achieve this by determining the sensitivity of the rate, chemical yield, and selectivity of reactions using selected molecular descriptors in a principle component analysis and then through a Design of Experiment (DoE) methodology. The resulting statistical model will be used to generate initial best-guess starting points for synthetic conditions. The model will be updated regularly with additional screening data for reagents and reaction conditions generated from automated optimisations. This feedback loop between statistical models and experimental results will increase in accuracy with each iteration, leading to a predictive tool for mapping reactivity with starting materials, catalysts and solvents external to the initial model. This structure-property/reactivity relationship combined with a statistical response surfaces (DoE) -based approach will then be integrated with a digital twin of the process, enabling a holistic approach to the optimisation of the reaction class, including the development of a decision- support framework.