Data-driven prediction of eutectic MgCl2–KCl molten salt viscosity using advanced machine learning across 553–1346 K

Eutectic MgCl 2 –KCl (32:68 mol%) molten salt has attracted considerable interest in high-temperature energy applications, including thermal-energy storage and heat-transfer systems. Accurate viscosity prediction is important for optimal system design and operation. In this study, 1045 experimental viscosity data were compiled from the literature, cleaned and screened, resulting in 987 retained observations. Seven conventional viscosity correlations were first evaluated, followed by development of a comprehensive machine-learning framework for modelling viscosity as a function of temperature (553–1346 K) and density (1.46–2.12 g cm -3 ). The models developed in this study included multilayer perceptron (MLP), cascade-forward neural network (CFNN), Group Method of Data Handling (GMDH) and hard-monotonic physics-informed MLP (PI-MLP), with temperature-only and temperature–density inputs. The data-driven neural networks were trained using a range of gradient-based and evolutionary optimisation algorithms. A five-fold cross-validation (CV) process was applied to an 80% development set to select the optimal model, while the remaining 20% was reserved as an independent test set. The MLP-SCG (2-25-25-1) model achieved the lowest CV root mean square error, with test and total RMSE values of 0.779 and 0.603 cP, respectively. In contrast, dependency-aware leave-one-source-out (LOSO) validation indicated reduced generalisation to previously unseen literature sources. MLP-Adam and MLP-BR demonstrated greater robustness at the source level, underscoring the significance of source dependency in evaluating model transferability. The hard-monotonic PI-MLP models produced physically consistent, non-negative apparent activation energies in the liquid domain, with median values of 5.49 and 8.34 kJ mol -1 for the temperature-only and temperature–density formulations, respectively. Williams-plot analysis indicated that 93.91% of test observations were within the applicability domain of MLP-SCG. The proposed framework provides an efficient means of estimating MgCl 2 –KCl viscosity while also allowing predictive accuracy, source-level generalisation, and physical consistency to be assessed separately.