Learned Response-Field Inertia Operator for HEC-RAS 2D Water-Surface Elevation Prediction

Problem
This work addresses the gap in predictive modeling for water-surface elevation (WSE) in HEC-RAS 2D simulations, specifically focusing on the limitations of existing surrogate models that often rely on raster remapping, which can introduce errors and confound information access. The authors present a cross-dataset evaluation of learned native-cell surrogate models, emphasizing the need for a method that can operate directly on nonuniform computational cells without the drawbacks of traditional approaches. This paper is a preprint and has not undergone peer review.

Method
The core contribution is the introduction of the Learned Response-Field Inertia Operator (LRFIO), which is a no-forcing, increment-based learned surrogate model. LRFIO calibrates an inertial response operator derived from solved HEC-RAS trajectories and implements this operator through a closed-form native-cell rollout. The model evaluates a hierarchy of responses, including persistence, global calibrated inertia, and segmented response-field inertia. The authors treat segmentation, residual correction, and neuralized inertia as learnable modeling choices, allowing for adaptive complexity based on validation evidence. The training process involves multiple HEC-RAS 2D benchmarks, ensuring that the model retains different response structures tailored to specific domains.

Results
LRFIO was evaluated across four diverse HEC-RAS 2D benchmarks, demonstrating a maximum validation regret of 4.30%, indicating controlled complexity. The deployment times for the retained rollout ranged from 0.003 seconds to 0.242 seconds. Notably, in a comparison with the Beaver Bayou measured-solve, LRFIO achieved an estimated speedup of 2.75 x 10^4 when normalized to the horizon, significantly outperforming traditional HEC-RAS simulations.

Limitations
The authors acknowledge that while LRFIO shows promise, its performance is contingent on the empirical justification for added complexity in response-field, neural, or spatial dimensions. They also note that the model’s effectiveness may vary across different hydrological contexts, which could limit its generalizability. Additionally, as a preprint, the findings have not yet been validated through peer review, which may affect their robustness.

Why it matters
The implications of this work are significant for the field of hydraulic modeling and simulation, as it provides a novel approach to WSE prediction that minimizes computational overhead while maintaining accuracy. The adaptive complexity of LRFIO allows for tailored modeling strategies that can enhance predictive performance in various hydrological scenarios. This research contributes to the ongoing development of efficient surrogate models in hydraulic engineering, as discussed in related literature on surrogate modeling techniques. For further details, see the full paper available on arXiv.