We are working with a company who, through their bespoke experimental work, have identified a number of highly promising material families that could transform power transmission, transport and computing. In order to accelerate their R&D efforts and stay ahead of the competition they wanted to explore and implement novel opportunities for machine learning that would compliment their material discovery workflows.
We undertook a comprehensive assessment of promising machine learning approaches in academic papers. We then built and tested a number of these models using our best data science practices to determine their real life utility for the client. Finally we developed an alternative staggered machine learning process that makes best use of the available data and can be integrated into their current activities.
We were able to identify and correct a common set of weaknesses in model design and validation that renders the academic approaches unsuitable for real-life application. The new process we created will help them identify and fast-track high priority materials for further experimental work and we continue to explore opportunities in this space as part of a longer term collaboration effort.