Machine-learning-based surrogate modeling with global sensitivity analysis and particle swarm optimization for pool boiling of ternary hybrid nanofluids

Pool boiling of R141b-based ternary hybrid nano-refrigerants was investigated experimentally and analyzed using an integrated framework combining artificial neural network (ANN) surrogate modeling, Shapley additive explanations (SHAP), Sobol global sensitivity analysis, and particle swarm optimization (PSO). The study used 358 measurements for eleven formulations over 1–3 bar and 3.44–269.6 kW m −2 . Increasing pressure raised the maximum measured boiling heat transfer coefficient ( h ) of pure R141b from 9.73 to 14.1 kW m −2 K −1 . The formulation containing 0.10 vol% SDBS and 0.10 vol% MWCNT reached the highest measured h , 14.7 kW m −2 K −1 at 3 bar, and retained 99.3% of this maximum at its highest measured heat flux. In contrast, the higher-loading ternary formulation remained 22.8–31.6% below the lower-loading ternary formulation. At 3 bar, the maximum h of the R141b + 0.10 vol% SDBS reference was 18.0% below that of pure R141b, demonstrating the importance of a matched surfactant baseline when interpreting nanoparticle effects. Screening 234 architecture–batch-size combinations on validation data selected a 12-linear/33-tanh/2-linear network with mini-batch size 4; retrained on a partition seed representative of eleven repeated splits, it reached test R 2 values of 0.889 for h and 0.920 for average boiling surface temperature ( T ₛ). SHAP and Sobol identified heat flux as dominant for h , while while SHAP identified pressure as dominant for T ₛ. Closure-constrained PSO predicted h = 14.9 kW m −2 K −1 at 3 bar and 269.6 kW m −2 for 99.8 vol% R141b, 0.10 vol% SDBS, and 0.10 vol% MWCNT. This boundary-located candidate requires experimental verification.