Prediction of delta-ferrite formation in 17-4 PH steel using zero-inflated Gaussian process

Alloy composition plays a pivotal role in the formation of phases and precipitates. A compelling example of this phenomenon is evident in the case of 17-4 PH steel, where even slight deviations in alloy composition can notably influence the formation of phases such as δ-ferrite. While thermodynamic calculations are invaluable in predicting the formation of phases, the procedure can become rather time-intensive, particularly when dealing with a wide array of alloy compositions. This issue is further exacerbated by the complex non-linear relationships that exist between the constituent elements and phases. While data-driven analyses can help, traditional machine learning (ML) models struggle to accurately predict δ-ferrite formation due to its zero-inflated nature. This is a result of how combinations of elemental compositions can either result in the presence or absence of δ-ferrite. This paper proposes a zero-inflated Gaussian process (GP) model, ZIGP, to understand the effect of alloying elements on the formation of δ-ferrite. This hybrid model combines a classifier and a regressor, showing an exceptional mean root relative squared error value of 0.407 and 0.843 for the prediction of the maximum δ-ferrite fraction and the δ-ferrite formation temperature, respectively. The ZIGP model is also used to understand the sensitivities of the δ-ferrite phase formation to different constituent elements. The model outperforms other traditional ML models applied to the zero-inflated data. The results of this research, thus, not only provide an understanding of the phases but can also be extended to optimize material composition and design appropriate heat treatment cycles to achieve desired properties.