Learned Response-Field Inertia Operator

• LRFIO is a surrogate model that predicts future water-surface elevations in HEC-RAS using native nonuniform computational cells without raster remapping.
• It employs an inertial response operator by calibrating current increments with cell-specific parameters and segmentation for enhanced accuracy.
• Empirical validation demonstrates LRFIO’s significant speedup and low error compared to traditional and neural network-based surrogates.

Searching arXiv for the cited LRFIO paper and related inertia/operator work to ground the article. The Learned Response-Field Inertia Operator (LRFIO) is a native-cell surrogate for predicting future HEC-RAS 2D water-surface elevations (WSE) directly on the solver’s original nonuniform computational cells. It was introduced for cross-dataset evaluation of solver-consistent, no-forcing WSE prediction, with the explicit aim of avoiding raster remapping error and information-access confounding. In LRFIO, the current native-cell increment is treated as the central predictive scaffold, and future WSE is generated by calibrating an inertial response operator from solved HEC-RAS trajectories and deploying the retained operator through closed-form native-cell rollout (Holmberg et al., 4 Jun 2026).

1. Definition and problem setting

LRFIO is defined in a setting where each HEC-RAS 2D project has its own nonuniform mesh of computational cells. The state for dataset $d$ is represented as

$\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$

where $N_d$ is the number of native HEC-RAS 2D cells and each column corresponds to one native cell. All predictions and errors are computed on these native cells, with no rasterization. This native-cell evaluation is intended to keep the surrogate’s support exactly aligned with the solver and to avoid remapping errors introduced by native-to-raster-to-native conversion (Holmberg et al., 4 Jun 2026).

The motivating problem is twofold. First, many recent surrogates operate on rasters rather than the irregular native cells, which introduces support mismatch, many-to-one aggregation, and non-negligible numerical error before learning begins; Beaver Bayou is reported to exhibit approximately $1$ m RMSE in a native-to-raster-to-native round trip. Second, many models use full-horizon future forcing as inputs, which changes the information-access regime relative to a no-forcing current-state surrogate. LRFIO is framed explicitly to separate these issues: it predicts future WSE fields on the original HEC-RAS cells using no future forcing and no future solver outputs, only the current and previous native-cell WSE and the current increment (Holmberg et al., 4 Jun 2026).

The central observation is that the current WSE increment,

$$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$

is a solver-conditioned response signal. In the formulation used by LRFIO, this increment is already the net effect of terrain, storage, roughness, boundary conditions, and numerical controls as processed by HEC-RAS up to time $t$. Future WSE is reconstructed incrementally as

$$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$

2. Formal operator structure

At forecast time $t$, LRFIO learns an inertial response operator that maps the current increment $\Delta \mathbf{W}_t$ to future increments $\widehat{\Delta \mathbf{W}_{t+k}}$. The operator has two structural components: a cell-to-response-field assignment $\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$0, and for each response-field bin $\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$1, a parameter pair

$\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$2

where $\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$3 is an inertia coefficient controlling persistence or decay and $\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$4 is an increment cap learned from data (Holmberg et al., 4 Jun 2026).

For cell $\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$5 in bin $\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$6, the propagated future increment at lead $\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$7 is

$\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$8

The corresponding WSE rollout is

$\mathbf{W}^{(d)} \in \mathbb{R}^{T_d \times N_d},$9

This defines LRFIO as a learned mapping that propagates the current increment field by inertial decay with capping, optionally segmented across response fields. The parameter bundle is written as

$N_d$0

where $N_d$1 records the retained response case, $N_d$2 is the segmentation, and $N_d$3 stores the selector audit. Calibration is performed offline using only training and validation HEC-RAS trajectories. At deployment, no additional optimization is performed; the retained operator is applied directly in closed form (Holmberg et al., 4 Jun 2026).

A common misconception is to treat LRFIO as a forcing-aware hydrodynamic predictor. It is not formulated that way. It is a no-forcing current-state surrogate whose deployment inputs are

$N_d$4

with no future forcing time series and no future solver outputs.

3. Response hierarchy and model family

LRFIO is embedded in a base-case-first response hierarchy consisting of three cases: $N_d$5 persistence, $N_d$6 global calibrated inertia, and $N_d$7 segmented response-field inertia (Holmberg et al., 4 Jun 2026).

In $N_d$8, future increments are set to zero,

$N_d$9

so that future WSE is simply held at the current value. This is mathematically equivalent to $1$0. The paper reports that persistence is retained for the Lower San Saba River dataset, where stage changes over the forecast horizon are tiny and the stage RMSE is $1$1 m. In that case, added inertia structure does not improve validation sufficiently to justify complexity.

In $1$2, the entire domain is treated as a single response region with shared parameters: $1$3 The update becomes

$1$4

This global inertia case is retained for Upper San Saba River, with $1$5 and $1$6 m, and for Tuttle Creek / Big Blue / Kansas River, with $1$7 and $1$8 m.

In $1$9, cells are assigned to multiple response fields according to a response score. A key score is the calibration WSE range,

$$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$0

Cells are then grouped into $$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$1 bins by quantiles of $$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$2. For each bin $$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$3, the increment cap is estimated from the calibration increment set

$$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$4

via a quantile $$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$5, while $$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$6 is chosen from a discrete grid including $$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$7. Segmentation is retained for Beaver Bayou, where two response bins give a substantial validation improvement over the global case. This suggests that LRFIO treats spatial heterogeneity as an empirical modeling choice rather than as a fixed prior.

The paper also evaluates two extensions: residual correction, denoted LRFIO-R, in which a ridge regression residual map is added after the closed-form rollout, and neuralized inertia, denoted INO, in which an MLP residual head is added to the inertia scaffold. These variants are treated as optional modeling choices rather than as part of the core definition (Holmberg et al., 4 Jun 2026).

4. Information-access policy, calibration, and complexity control

A defining feature of the framework is an explicit information-access policy separating static project inputs, current hydraulic state, project-input forcing, calibration-derived quantities, and future solver-output targets. For LRFIO, deployment is restricted to current and previous WSE, the current increment, and the calibrated parameter bundle. Future forcing and future solver outputs are excluded from the input set (Holmberg et al., 4 Jun 2026).

Calibration proceeds per dataset over a candidate family

$$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$8

For each candidate, the method constructs segmentation if applicable, estimates caps, chooses inertia coefficients from small discrete grids, rolls out predictions on validation windows, and computes a scalar validation score $$\Delta \mathbf{W}^{(d)}_t = \mathbf{W}^{(d)}_t-\mathbf{W}^{(d)}_{t-1},$$9. The selection score is

$t$0

Here $t$1 is overall stage RMSE, $t$2 is final lead-time RMSE, $t$3 is the $t$4th percentile absolute error at final lead, $t$5 is hotspot RMSE on the $t$6 cells with the largest final-lead error, and $t$7 is overall stage bias.

Complexity control is structural rather than parametric. There is no explicit $t$8 penalty in the closed-form LRFIO itself; the regularization mechanism is model-family selection together with grid-based parameterization. The selector is based on validation regret

$t$9

and segmented gain

$$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$0

The final rule retains the best base case when $$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$1, retains the best segmented case only when $$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$2 and $$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$3, and otherwise reverts to the best base case, with $$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$4 and $$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$5. This enforces controlled complexity with bounded regret (Holmberg et al., 4 Jun 2026).

5. Evaluation across HEC-RAS benchmarks

The reported evaluation uses four diverse HEC-RAS 2D projects with chronological splitting and per-project calibration (Holmberg et al., 4 Jun 2026).

Dataset	Native cells and horizon	Retained case
Beaver Bayou	9,236 cells; 12 h (24 steps)	$$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$6 segmented inertia
Upper San Saba River	430,874 cells; 12 h (48 steps)	$$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$7 global inertia
Lower San Saba River	201,676 cells; 12 h (48 steps)	$$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$8 persistence
Tuttle Creek / Big Blue / Kansas River	44,255 cells; 6.1667 h (37 steps)	$$\widehat{\mathbf{W}_{t_0+k}} = \mathbf{W}_{t_0} + \sum_{j=1}^{k} \widehat{\Delta\mathbf{W}_{t_0+j}}, \qquad k=1,\ldots,H.$$9 global inertia

The broader benchmark includes raster CNNs and UNets, graph GNNs including SWE-GNN, Fourier Neural Operators, DeepONet, recurrent models, reduced-order models, and forcing-aware variants. Models are not trained jointly across datasets; rather, the framework and selection logic are evaluated across distinct domains.

Empirically, the retained response structure differs by domain. In three of four datasets, the retained model is also the absolute-best candidate by validation score. Tuttle Creek is the main exception: the absolute-best candidate is a $t$0-bin segmented model, but its segmented gain over the best base case is approximately $t$1, below the $t$2 threshold, so the selector retains the simpler global inertia model and accepts approximately $t$3 regret. Across all datasets, the mean regret is approximately $t$4 and the maximum regret is $t$5. The selector audit is therefore presented as evidence of parsimony rather than automatic escalation to the most complex model (Holmberg et al., 4 Jun 2026).

6. Empirical performance, interpretation, and scope

The speed-accuracy results place the LRFIO family close to the Pareto frontier of low error and low runtime across the four benchmark domains. For Beaver Bayou, a representative LRFIO achieves stage RMSE of approximately $t$6 m with rollout time approximately $t$7 s, while the best non-LRFIO comparator, an FNO, has RMSE approximately $t$8 m and runtime approximately $t$9 s. For Upper San Saba, LRFIO and SWE-GNN tie at RMSE approximately $\Delta \mathbf{W}_t$0 m, with SWE-GNN slightly faster but of the same order. For Lower San Saba, LRFIO achieves RMSE approximately $\Delta \mathbf{W}_t$1 m with runtime approximately $\Delta \mathbf{W}_t$2 s, versus SWE-GNN at approximately $\Delta \mathbf{W}_t$3 m and $\Delta \mathbf{W}_t$4 s. For Tuttle Creek, LRFIO has RMSE approximately $\Delta \mathbf{W}_t$5 m and runtime approximately $\Delta \mathbf{W}_t$6 s, while DeepONet has RMSE approximately $\Delta \mathbf{W}_t$7 m and runtime approximately $\Delta \mathbf{W}_t$8 s (Holmberg et al., 4 Jun 2026).

Retained rollout times range from $\Delta \mathbf{W}_t$9 s for Beaver Bayou to $\widehat{\Delta \mathbf{W}_{t+k}}$0 s for Upper San Saba. For Beaver Bayou, a measured full $\widehat{\Delta \mathbf{W}_{t+k}}$1 h HEC-RAS simulation takes $\widehat{\Delta \mathbf{W}_{t+k}}$2 s; horizon-normalized to the $\widehat{\Delta \mathbf{W}_{t+k}}$3 h forecast horizon, this yields approximately $\widehat{\Delta \mathbf{W}_{t+k}}$4 s, compared with $\widehat{\Delta \mathbf{W}_{t+k}}$5 s for LRFIO, corresponding to an estimated horizon-normalized speedup of approximately $\widehat{\Delta \mathbf{W}_{t+k}}$6.

The central interpretation is that the current native-cell increment is a strong predictive scaffold. In the reported experiments, added response-field, neural, or spatial complexity is retained only when validation evidence justifies its cost. Residual correction is always worse than base LRFIO in both runtime and RMSE across all four datasets. INO yields a small RMSE improvement on Upper San Saba, but at approximately $\widehat{\Delta \mathbf{W}_{t+k}}$7 higher runtime, and on the other three datasets it is both slower and less accurate. A plausible implication is that, under a matched no-forcing information policy, much of the apparent advantage of more elaborate surrogate families can disappear when the native-cell increment is preserved and propagated directly (Holmberg et al., 4 Jun 2026).

LRFIO remains limited in specific ways. It is calibrated per project rather than with cross-domain parameter sharing; it models only WSE; it does not explicitly incorporate future rainfall or boundary hydrographs; and the reported horizons are $\widehat{\Delta \mathbf{W}_{t+k}}$8 to $\widehat{\Delta \mathbf{W}_{t+k}}$9 hours. It is also solver-consistent rather than observation-consistent: its target is the HEC-RAS solution, so errors in the underlying HEC-RAS configuration are inherited by the surrogate.

The term “inertia” in LRFIO should not be conflated with its use in nonlinear oscillator identification. In “response-dependent inertia” for forced oscillatory systems, the problem is the identification of instantaneous modal parameters when generalized inertia depends on measured response dynamics (Harduf et al., 2023). LRFIO instead uses “inertia” to denote a learned increment-propagation mechanism for future WSE on native HEC-RAS cells. This suggests a shared vocabulary of response propagation rather than a shared methodology.