Optimal design of steel grades to meet a target hardenability profile through evolutionary computation and autoencoders

Colla Valentina - Scuola Superiore Sant'Anna, TeCIP Institute (Italy)

Hardenability is one of the most important mechanical features of steel for automotive applications and for many other different uses. This property, that represents the steel capability to improve its hardness through heat treatment, is usually assessed via the Jominy end-quench test, that outputs a curve of hardness values measured in different positions of a specimen quenched and tempered according to a standardized procedure. The Jominy curve shape is affected by the content of Carbon and other micro-alloying elements, e.g. Chromium, Manganese, Molybdenum, Silicon, Nickel, and Boron, but the interactions among micro-alloying elements are not yet completely understood.

To ensure the steel suitability for specific usages, most customers of steel companies define constraints and/or target values for several points of this hardenability profile. The problem, therefore, arises to determine or design the most appropriate steel grade to meet a given target, and the solution might not be unique. Moreover, while customers’ requirements are becoming increasingly challenging, the steel sector is pressurized by ever more stringent environmental regulations and relevant costs of ferroalloys. To sum up, in the design of a steel grade meeting a target hardenability profile, it is important to consider not only the end-user’s constraints but also other factors, such as cost, availability and environmental impact of micro-alloying elements.

The paper presents an optimization approach based on evolutionary computation, which exploits a model using autoencoders (a type of artificial neural network) to estimate the Jominy profile of a steel from its chemical composition. The proposed can be customized to the production range and targets of any steel company, in terms of both datasets exploited for internal model training, and optimization targets. The flexible formulation of the objective function of the optimization problem allows jointly considering several objectives by weighing them based on their importance for the company’s strategy.