(362l) Cost Optimization of Steel Alloying Elements with Machine Learning Based Jominy Hardness Profile Prediction

Prediction of relevant mechanical properties is vital for the steel industry and for materials engineering. This work investigates a novel method to predict the Jominy hardness curve of a single grade of steel based solely on the composition of the alloying elements. Further consideration is given to the optimisation of the chemical profile of the alloying elements with respect to the overall cost. The ultimate aim of the work is to be able to reduce spend on alloying elements by offering cheaper methods of achieving the same Jominy profile based upon accurate prediction. Use of a feature auto-encoder neural network is employed to reduce the dimensionality of 19 alloying elements from some 1000 steel samples to extract relevant property information. Gaussian Process (GP) regression is used to predict the Jominy hardness profile of the steel up to a depth of 10mm. Results showed that a GP fed with extracted features was able to predict Jominy hardness profiles with high accuracy (RMSE_unseen = 1.35, MAPE = 3.3%). Constrained optimisation using the GP model, and historical market data for the price of alloying elements showed that an average cost reduction of 18% on alloying element spend is possible. This work shows not only that novel machine learning approaches can be used to accurately predict mechanical properties of a given grade of steel, but crucially that use of these in conjunction with optimisation can generate considerable financial saving.