Chemistry, particularly synthetic chemistry, is undergoing a profound transformation in its practices with the advent of robotics, flow chemistry, in-line monitoring and digital tools. Optimization is one of the areas of synthetic chemistry most affected by the digitalization of chemistry because, by nature, it aims to improve existing practices and processes by any means necessary, including the use of new technologies such as robotics, flow reactors, in-line monitoring and artificial intelligence (AI). Virtually every synthetic chemist can be involved in optimization processes, whether in R&D for improving a methodology or synthesis, or in process development for setting up a manufacturing process. However, shifting to digital chemistry is not as straightforward as hoped, as it faces limited available expertise, and a growing demand for data to train AI models.
The ChemAI project aims at transforming the efficiency of AI-assisted chemical reaction optimization by i/ requiring fewer experiments, ii/ delivering faster results and iii/ acquiring better-quality data. This will be achieved through the development of innovative solutions based on an original perspective. Specifically, we will:
- Minimize the amount of experimental data required by employing transfer learning to create a digital twin of human intuition and by developing constrained optimization strategies that prevent futile experiments.
- Accelerate data analysis through the development of transient flow NMR methods at high field
- Achieve quantification within complex mixture by developing advanced NMR pulse sequences and modelling strategies
- Apply these methodologies to develop and optimize new metal-catalyzed oxidative couplings.
These novel methods will make AI-assisted optimisation (of yield, productivity, cost…) more broadly accessible and applicable for synthetic and process chemists of all backgrounds. To achieve this, the project will rely on two pillars: Bayesian algorithms and high-field NMR spectroscopy. ChemAI will lead to more efficient Bayesian approaches by integrating human and bibliographic knowledge transfer and developing constraints that enable more relevant experiment suggestions. It will also speed up autonomous training and optimization with transient flow methods. These methodological developments will be integrated into a flow optimization platform for the optimization of chemical reactions.
