Abstract:

Physical metallurgical (PM) and data-driven approaches can be independently applied to alloy design. Steel technology is a field of physical metallurgy around which some of the most comprehensive understanding has been developed, with vast models on the relationship between composition, processing, microstructure and properties. They have been applied to the design of new steel alloys in the pursuit of grades of improved properties. With the advent of rapid computing and low-cost data storage, a wealth of data has become available to a suite of modelling techniques referred to as machine learning (ML). ML is being emergingly applied in materials discovery while it requires data mining with its adoption being limited by insufficient high-quality datasets, often leading to unrealistic materials design predictions outside the boundaries of the intended properties. It is therefore required to appraise the strength and weaknesses of PM and ML approach, to assess the real design power of each towards designing novel steel grades. This work incorporates models and datasets from well-established literature on marageing steels. Combining genetic algorithm (GA) with PM models to optimise the parameters adopted for each dataset to maximise the prediction accuracy of PM models, and the results were compared with ML models. The results indicate that PM approaches provide a clearer picture of the overall composition-microstructure-properties relationship but are highly sensitive to the alloy system and hence lack on exploration ability of new domains. ML conversely provides little explicit physical insight whilst yielding a stronger prediction accuracy for large-scale data. Hybrid PM/ML approaches provide solutions maximising accuracy, while leading to a clearer physical picture and the desired properties.

Key words: Machine learning, Physical metallurgy, Small sample problem, Marageing steel