CALPHAD-Based Multi-Objective Composition Screening of P91-Type Steel with Explainable Machine Learning

The thermodynamic competition among multiple precipitate phases and its coupling with solidification behavior in P91-type steel remain insufficiently explored across multicomponent composition spaces, limiting the systematic identification of favorable alloy compositions. To address this gap, a framework integrating high-throughput CALPHAD calculations, explainable machine learning, and multi-objective screening was developed. A full-factorial design of 26,244 compositions yielded 19,838 complete datasets after excluding invalid calculations. The equilibrium mole fractions of M23C6, MX, AlN, Z phase, and Laves phase at 600 °C, together with the Scheil solidification interval, were investigated. Random Forest, XGBoost, and LightGBM models were evaluated, and SHAP analysis was used to interpret the composition-dependent predictions of thermodynamic responses. C was identified as the dominant contributor to the calculated M23C6 fraction and to the Scheil solidification interval, whereas N, Nb, and V strongly influenced the calculated MX fraction. Within the investigated composition space, increasing N was associated with a lower equilibrium MX fraction but higher AlN and Z-phase fractions, indicating competition among N-containing precipitates under the adopted equilibrium conditions. Direct Pareto screening of the CALPHAD results identified 812 non-dominated compositions. Subsequent TOPSIS evaluation favored C = 0.12, Mo = 0.85, Nb = 0.10, and N = 0.03 wt.% under the selected criteria. Relative to the base P91 composition, the highest-ranked candidate exhibited reductions of 39.99% and 98.30% in the Z-phase and Laves-phase fractions, respectively, together with a 12.93% increase in the M23C6 fraction. However, its Scheil solidification interval increased by 23.36 °C, revealing a trade-off between precipitation-related objectives and solidification behavior. The selected composition therefore represents a thermodynamically favorable candidate within the investigated discrete design space and under the adopted screening criteria, rather than a universally optimal composition for P91 steel.