Solving Scale Up in the Chemical Industries: Identifying the Biggest Challenges and the Impact of Machine Learning

Machine learning has changed the paradigm of laboratory chemistry, with a large drive into using algorithms to make decisions, optimize, and analyse experimental results. The largest impact of this has been made for the discovery of new materials, drugs, and agrochemicals with billion-valuation companies emerging to dominate the scene such as Isomorphic Labs and Cusp AI. However, the scale up and process development of new products has been left behind by AI for Science. At SOLVE Chemistry, we are looking to bring the benefits of machine learning to scale-up science. Driven by the idea that true impact is blocked by a lack of data, we have built an automated laboratory powered by a stack of novel machine learning algorithms with the long-term goal of first-time right scale up, bringing more new chemicals to market, more sustainably, and less expensive.

We will first briefly explain transient flow, an emerging methodology that allows us to efficiently gather data by slowly varying conditions of a reactor and constantly measuring the stream of data. While the technique allows us to be very efficient at collecting data, it poses new challenges for Bayesian optimisation and active learning in such reactors, which necessitated the development of transition-constrained methods that can plan ahead whole reactor campaigns, acquiring the best data in an order that minimizes the amount of steady-state waiting times. We will showcase the algorithm in practice, used in a solvent replacement campaign, and explain further developments and challenges that arise. We will then look at how Large Language Models can be used to generate insights after the reaction data is collected, creating not an optimization loop, but a scientific discovery loop where the target becomes the discovery of true underlying chemistry that is scale-independent and can be used to speed up process development.