Speaker
Description
This study develops supervised machine-learning surrogate models for the molar volume of compositions within the Ni–Co–Cr ternary system, trained on CALPHAD-generated data. Molar volume is a fundamental thermodynamic quantity related to density and to temperature-dependent volumetric expansion, and is therefore relevant to the design of compositionally graded alloys. The original dataset contained 10,000 composition–temperature records spanning 300–2200 K. After the ternary closure condition was enforced, 4,998 physically consistent records were retained. To eliminate exact multicollinearity, the Ni and Co atomic fractions and the temperature were used as independent inputs, and Cr was reconstructed from the closure relation. Seven regression models were evaluated on a common 80/20 train–test split and ranked by held-out root-mean-square error (RMSE). Random Forest was the most accurate, with a test RMSE of 0.0291 cm3/mol and R2 = 0.9963, followed by Gradient Boosting (0.0540 cm3/mol, R2 = 0.9872) and support vector regression (0.0738 cm3/mol, R2 = 0.9762). Fourth-order polynomial regression remained competitive (0.0884 cm3/mol, R2 = 0.9658) while offering the clearest analytical representation and the lowest fitting time. The results show that machine-learning models can reproduce CALPHAD outputs efficiently within the sampled domain. Experimental validation is required before quantitative engineering use.