Tsunami-Surrogate Benchmark: Processed Multi-Reference Dataset for Neural-Operator Emulation of Tsunami-Like Wave Propagation

This record contains the Common-Time V2 reproduction package accompanying the manuscript Reference-Aware Evaluation of Neural PDE Surrogates: A Controlled Multi- Reference Benchmark for Tsunami-Like Waves. The benchmark contains 13,500 shared synthetic bathymetry–source scenarios evaluated using three numerical references: Hydrostatic, MUSCL-HR, and a linear weakly dispersive Boussinesq model. Each reference retains its natural internal time stepping while publishing 50 outputs at the same requested times from 0.0035 to 0.175 benchmark-time units. Numerical computation uses a buffered 96 × 96 domain with a central 64 × 64 publication crop. The package includes: - Complete processed train, validation, and test datasets for all three references. - Selected model checkpoints, resolved configurations, and training histories. - The validated final evaluation run and numerical-evidence summaries. - Strict-family holdout, native-resolution MUSCL-HR, and real-bathymetry evaluation data. - Reproducible plotting and statistical-analysis scripts. - Elsevier manuscript sources, generated figures, and the compiled paper PDF. - SHA-256 checksums and manifests describing every archive. The package supports reproduction of the reported model-accuracy, solver-gap, cross-reference, uncertainty, resolution-transfer, morphology-transfer, and numerical-verification analyses. The main processed containers hold Common-Time V2 trajectories inside an older outer manifest envelope; the final evaluation audit found no legacy saved-step trajectories in the reported evaluation data. This is a controlled, synthetic, finite-horizon research benchmark. It is not an operational tsunami-warning, navigation, inundation, run-up, or site-specific hazard product. The numerical references are explicit label generators rather than physical ground truth, and the reported GeoClaw/Clawpack comparisons do not establish pointwise production equivalence or event-scale physical validation. Source code and documentation are available from the project repository. File integrity can be verified using the included SHA256SUMS.txt. DOI: 10.5281/zenodo.21956834