Supervised learning to support the optimisation of chemical reactions

Jean Golding Institute Seedcorn Funding, £4974.


Chemical experimental data provides many challenges for data scientists. Firstly, the approach to experimental design and screening has traditionally been driven by feasibility, leading to significant inefficiencies if the optimum areas of input space are missed. Where time and resources are scarce, the sample size of experiments can also be relatively small, whilst the response surfaces can vary drastically in relation to small changes in inputs; robotic screening is helping to address this, but new data analysis approaches are needed to optimally use the resulting data. In contrast, statistical learning of the underlying process aims to produce parsimonious and chemically interpretable models, which have the statistical power to identify new directions of interest, making experimental screening more efficient and more effective.

Data analysis must take these challenges into consideration and we plan to explore different methodologies, with a view to informing future analyses, as well as engaging with the chemical and statistical communities to identify necessary developments and then preparing applications for follow-on funding.

The PI holds experimental data collected as part of a case study during her EPSRC fellowship (2007-2012); we will seek to re-analyse these results with more advanced statistical approaches. In addition, we plan to engage with academic and industrial specialists in reaction screening and optimisation, with the view to applying our insights to more multivariate results.