Date: 30 October 2025 | Category: News
Authors: Perman Jorayev, Sebastian Soritz, Simon Sung, Mohammed I. Jeraal, Danilo Russo, Alexandre Barthelme, Frédéric C. Toussaint, Matthew J. Gaunt, and Alexei A. Lapkin*
The Lapkin group at the University of Cambridge, in conjunction with UCB Pharma, have described a machine learning-driven optimisation of photoredox tertiary amine synthesis, under a semiautomated continuous flow set-up, with six continuous variables and solvent choice as a discrete variable.[1] Yield and reaction cost were key objectives, with a novel Bayesian optimization algorithm, nomadic exploratory multiobjective optimization (NEMO), deployed to identify and populate the Pareto front. It was found that catalyst loading and residence time were correlated to absorbed photon equivalence, with catalyst loading the main contributor to the overall cost. Outputs using flow chemistry were 25 times higher than in batch, equating to a 12 g/day throughput.
Figure 1: General overview of the photoredox amine synthesis
Photoredox catalysis: the renaissance
While photoredox catalysis was developed over 40 years ago,[2] in recent years, visible light photocatalysis has experienced a renaissance, with rapid adoption by both academia and industry.[3, 4] This is likely tied to the advent of flow chemistry, which can offer significant advantage including uniform reaction irradiation, improved reaction scalability, improved reproducibility and improved mixing and heat exchange, often with substantially higher yields than achieved in batch. However, while use of flow chemistry has facilitated efficiency in terms of reaction time and reaction control, reaction optimisation and development of novel chemical transformations still requires significant time, effort and resources.
Machine learning and Bayesian optimisation for reaction design
Machine learning (ML), or more specifically Bayesian optimization (BO) algorithms, provide opportunity to exploit available data to build statistical models that can be used to optimise multiple, sometimes competing, process objectives (multiobjective optimisation). Multiobjective optimisation has been used in many “self-driving labs”, where experimental workflows and experimental planning are automated. [5]
Flow chemistry gives higher yields than batch
During this work, ML optimisation of amine synthesis was demonstrated using six continuous variables and 20 solvents under a semiautomated flow set-up. With an initial library of 115 solvents, the workflow included steps for gathering of a priori knowledge and incorporated physical observations such as identifying a subset of 20 solvents (with Hantzsch ester solubility being a key criterion), and UV-visible photo flux experiments providing a guideline for optimal reactor, lamp, and tubing selection. Once assembled, the NEMO algorithm, trained on experimental data generated using an eight solvent subset, was employed to identify and populate the Pareto front for the two reaction objectives – yield and cost – under semiautomated continuous flow conditions. This approach ultimately identified conditions that gave a productivity value of ∼12 g/day during scale-up when using a Vapourtec UV-150 reactor fitted with a 470 nm LED lamp, which was ∼25× higher than the optimal batch result.
References:
[1] Machine Learning-Driven Optimization of Continuous-Flow Photoredox Amine Synthesis (P. Jorayev, S. Soritz, S. Sung, M. I. Jeraal, D. Russo, A. Barthelme, F. C. Toussaint, M. J. Gaunt, A. A. Lapkin, Org. Proc. Res. Dev., 2025, 29, 1411–1422). https://doi.org/10.1021/acs.oprd.4c00533
[2] Chemistry of dihydropyridines. 9. Hydride transfer from 1,4-dihydropyridines to sp3-hybridized carbon in sulfonium salts and activated halides. Studies with NAD(P)H models (T. J. Van Bergen, D. M. Hedstrand, W. H. Kruizinga, R. M. Kellogg, J. Org. Chem., 1979, 44, 4953–4962). https://doi.org/10.1021/jo00394a044
[3] Visible Light Photoredox Catalysis with Transition Metal Complexes: Applications in Organic Synthesis (C. K. Prier, D. A. Rankic, D. W. C. MacMillan, Chem. Rev., 2013, 113, 5322–5363). https://doi.org/10.1021/cr300503r
[4] Photoredox Catalysis in Organic Chemistry (M. H. Shaw, J. Twilton, D. W. C. MacMillan, J. Org. Chem., 2016, 81, 6898–6929). https://doi.org/10.1021/acs.joc.6b01449
[5] Self-Driving Laboratories for Chemistry and Materials Science (G. Tom, S. P. Schmid, Sterling G. Baird, Y. Cao, K. Darvish, H. Hao, S. Lo, S. Pablo-García, E. M. Rajaonson, M. Skreta, N. Yoshikawa, S. Corapi, G. Deniz Akkoc, F. Strieth-Kalthoff, M. Seifrid, A. Aspuru-Guzik, Chem. Rev., 2024, 124, 9633–9732). https://doi.org/10.1021/acs.chemrev.4c00055
Learn more about the UV-150 photochemical reactor