(529b) Closed Loop Multi Target Optimization for Discovery of Chemical Reactions

Process development in the pharmaceutical industry involves extensive experimental work aimed at process optimisation and developing process understanding. Recent advances in automation and machine-learning based design of experiments aim to minimise the number of expensive experiments needed in process development and hence to free up researchers from routine work. This capability may also be used in discovering new materials and process recipes. The work
herein extends a framework in the rapidly growing area of closed loop optimisation (also known as self optimisation [2]), by incorporating a priori knowledge of the chemical system under investigation. The case study is a catalytic C-H activation reaction to prepare an aziridine product. A model-based design of experiments method was adopted to create experiments that deliver maximum information for the estimation of kinetic parameters. This used a physical model to describe the chemical phenomena and represent the a priori knowledge. These experiments were conducted in an automated small scale continuous flow reactor system and product concentration was measured on-line by gas chromatography. In addition to estimating kinetic parameters for the physical process model, the experimental results were then used to train a machine-learning based multi-objective statistical algorithm for target optimisation to attain a specific cost function value and maximise the yield of the process. This algorithm was further trained in silico using the process model until successful recipes with respect to the targets were proposed. These recipes were subsequently tested experimentally to confirm their promise. The results showed that even the first of these tests agreed well with the prediction of attained targets, validating the effectiveness of incorporating a priori knowledge in the optimisation framework. Finally, this approach was compared to one in which a priori knowledge was not exploited and the iterative optimisation process was driven by physical experiments exclusively, whereby the initial training set was derived using Latin hypercube sampling. This approach, too, succeeded in achieving target results in very few iterations. The use of machine-learning algorithms alongside development of mechanistic process models has not been developed so far, but this work shows that it offers a promising route to rapid generation of process knowledge in the case of complex chemical transformations.