Physics-prior-guided stress concentration factor prediction for marine engine crankshaft

To address the computational cost of finite element analysis and the small-sample boundary extrapolation distortion of data-driven surrogate models, an additive residual prediction framework (ARPF) integrating classification society empirical priors is proposed for predicting crankshaft transition fillet stress concentration factors (SCFs). The ARPF utilizes an analytical empirical formulation as a macroscopic baseline to explicitly de-trend the response space, compressing target physical residuals into a stationary, narrow interval and substantially reducing regression complexity. Five surrogate paradigms were evaluated across sample sizes from N = 25 to 250 against an independent test set. Under the small-sample regime of N = 50, direct prediction exhibited localized error spikes, with maximum relative errors (MaxRE) of random forest and support vector regression reaching 25.81% and 24.53%, respectively. In contrast, the ARPF improved the coefficient of determination (R 2 ) of all five models to above 0.996 and restricted MaxRE below 13%. Pairwise Wilcoxon signed-rank tests and non-parametric effect sizes across ten scrambled Sobol repetitions confirmed the robustness of these accuracy gains. Mechanistic analysis revealed that the prior-anchored macroscopic monotonic trend suppresses boundary extrapolation divergence, reliably supporting structural optimization with minimal finite element evaluations.