Data-Driven Reduction of Rail Pad Dynamic Stiffness Testing: Benchmarking a cGAN Against Gaussian Process Kriging

Full-factorial characterization of rail-pad dynamic stiffness is costly because each operating combination of toe load, load amplitude, and frequency must be tested under every experimental condition. This study evaluates whether data-driven reconstruction can support reduced testing campaigns by deliberately withholding operating combinations and benchmarking a grid-aware conditional generative adversarial network (cGAN) against Gaussian process (GP) kriging and other reconstruction methods. The hidden-cell protocol was applied to 360 measurements from an EPDM pad under ten experimental conditions and independently replicated on two additional pads. At the primary 25% masking level, the cGAN and pooled GP achieved mean repetition-level median absolute percentage errors of 4.8% and 1.0%, respectively, although unrestricted random masking left only 39 of 90 withheld observations interpolative on average. In S1-R, mask selection was conditioned a priori on preserving nominal geometric support, using only the nominal design coordinates and retaining all eight vertices. All 90 withheld observations were then evaluated at their achieved coordinates; the median across repetition-level median errors was 4.38% for the cGAN and 1.23% for pooled GP, with 80.89% and 99.78% of predictions within 10% error, respectively. The GP also provided better-calibrated uncertainty (95% coverage: 96.2% versus 84.8%), and its superior accuracy was reproduced on both additional pads. Overall, the results support the retrospective feasibility of a 25% reduction in tested operating combinations under a nominal support-preserving design, with pooled GP kriging providing the most accurate and robust reconstruction.