Value from corporate chemical data: Leveraging historic project data using machine learning

Large chemical organizations have run dozens of R&D projects over the years: but typically the data from these projects languishes in corporate repositories (or even spreadsheets - not to mention filing cabinets) without providing value to ongoing research. Machine learning and artificial intelligence promise to revolutionize the design, development, and processing of chemicals and formulations, but to do so they need to make the best use of all available data, ideally without being constrained to data from a single chemistry research project. These two problems form natural complements: using legacy data we can provide a firm machine learning-driven foundation for new projects, creating a virtuous cycle where each new project stands on the increasingly lofty shoulders of its forebears.

We will describe projects at Johnson Matthey that used data from past projects to build machine learning models that provided powerful insights into their design of new catalysts, and how this was enabled by a forward-thinking approach to handling experimental data. We will also discuss the gaps in current data storage and processing approaches and how these gaps can be plugged, generating the maximum value from existing data with a minimum of effort.