Fast Endpoint-Aware Residual Surrogate Modeling for Preliminary Static Screening of Deep-Water Catenary Mooring Lines

Within a single-depth, fixed line property, single-line synthetic quasi-static scope, this study develops an endpoint-prioritized two-stage training framework for full catenary geometry and terminal tension prediction. The global three-dimensional (3D) reference configuration first learns the coupled output and then applies validation-only selection for endpoint-focused refinement. An axisymmetry-informed radial plane representation enhancement is evaluated as the preferred accuracy configuration within the frozen seed-42 scope, yielding geometry root mean square error (RMSE) of 0.360 m, endpoint RMSE of 0.344 m, and tension mean absolute error (MAE) of 1641.8 N. Software-level cross-validation against MoorPy completed 1000 of 1000 requested cases, while a five-seed global-3D ensemble produced an empirical ensemble spread indicator whose geometry spread error Spearman correlation was 0.701. Runtime depended on batch size: central processing unit (CPU) batch-1 inference provided no advantage over the reference solver, whereas CPU batch-1000 throughput was 29.78× faster on the same machine. Expanded screening retained 19 unresolved solver failures. These results support preliminary static screening only. This software-level cross-validation does not constitute experimental, field, or external physical validation. The evidence does not establish dynamic, multiline, or calibrated uncertainty validation.