The new Sir Henry Royce Institute (SHRI) for Advanced Materials will provide £235m of capital investment over the next five years. One of nine core areas in the SHRI is Advanced Metals Processing. Computationally guided materials design is widely seen as a primary means of improving the efficiency and rate of discovery of new, improved materials. Key aspects of such projects are the generation (or accumulation) and subsequent management of large quantities of property and performance data. Significant progress has been made in these areas. The most challenging step in such projects is the successful use of such data - how are we to make useful predictions about new candidate materials? How are we to determine which materials might be the best candidates for some proposed application?
Machine learning and data analytics are finding wide application in materials chemistry and biology, but the field of metallurgy presents some particular challenges to the successful use of machine learning, for example:
1) The useful properties of an alloy are rarely dependent solely (and often not even mostly) on the properties of a perfect single crystal (or the basic unit cell). Rather it is the microstructure that plays the determining role.
2) The microstructure of a final, useable alloy is arrived at via an often complicated series of processing steps, with the evolution at different length and time scales being complex and interdependent.
3) Characterisation of the microstructure of materials is still far from automated and so is costly and time consuming. Raw data for machine learning are therefore expensive and coverage can be sparse.
4) Brute force, high accuracy simulation of microstructural features is, except in special cases, beyond current computational capabilities. Consequently we need to think not just about modelling methods, but also about the design of high-throughput small-scale processing and high-throughput experimental evaluation/screening.
5) We do not yet have an agreed basis for quantifying/representing a materials microstructure for input into machine learning methods.
6) We do not have the databases necessary to embark on machine learning approaches.