Ocean models need to know how fast internal gravity waves pass energy from large to small scales, because that flux sets how strongly the ocean mixes. The standard tool is the wave kinetic equation: an evolution law for the wave energy spectrum, the distribution of wave energy over wavenumber and frequency, rather than for individual wave crests. Its core is a collision integral, a sum over all resonant triads of waves that exchange energy. That integral is expensive to evaluate. Wu et al. made it feasible by reducing a 6-dimensional integral to a 2-dimensional one under horizontal isotropy. They found that "local" triads, in which all three waves have comparable scales, carry most of the downscale flux in the canonical Garrett–Munk (GM76) spectrum. Wu and Pan then showed a second effect for spectra steeper or shallower than GM76 ("red" or "blue"). There, a scale-separated mechanism called induced diffusion, in which a large-scale wave strains small-scale ones, acts as a one-way pipeline to dissipation that supplies up to 25% of the total. A fast learned approximation of the collision integral, a neural surrogate, would allow non-stationary spectra to be evolved and the integral to be embedded in larger models. Neural networks, however, have a known simplicity bias: they learn smooth, low-frequency structure first and sharp structure last (Rahaman et al.).
Our claim is that a single neural surrogate trained on the total collision integral will be biased in a physically specific way. It will learn the dominant, smooth local-interaction contribution early. It will under-learn the small, sharply structured induced-diffusion contribution. As a result it will look accurate on GM76, where induced diffusion is nearly zero, and silently drop a large share of dissipation on red and blue spectra. We propose to build the surrogate the way the physics is organized. One branch learns local interactions. A second branch learns induced diffusion and receives an extra input, the gradient of the action spectrum, which Wu et al. identify as the quantity whose vanishing makes induced-diffusion flux disappear in GM76. Supplying this input turns induced diffusion from a hard target into an easy one. This is the "priming" strategy of PrimeNet, applied to a physics surrogate rather than to a shortcut-prone classifier.
Core Claim
The proposal has three linked hypotheses. The first concerns a monolithic surrogate, meaning one network that maps an action spectrum to the full collision integral. Under a fixed training budget, its error on the induced-diffusion share of dissipation will be much larger than its error on total flux. This gap will grow as the spectral slope moves away from GM76, and plain loss curves will hide it. The second hypothesis is that the bias follows simplicity-bias ordering. If we project the surrogate's prediction onto the solver's interaction classes during training, the local-interaction component should converge first and induced diffusion last, mirroring the early-versus-late feature acquisition that Utama et al. and SEBRA observe for shortcut and hard features. The third hypothesis is that splitting the surrogate along the solver's own local versus scale-separated decomposition, and priming the scale-separated branch with spectral gradients, removes the bias. It should also extrapolate better to unseen spectral slopes.
The novelty is in how the two literatures are used. The debiasing papers turn simplicity bias into a detector of spurious image or text cues. We use it as a diagnostic and design principle for a physics surrogate, where the "hard" signal is a known, physically labelled mechanism. The wave-turbulence papers evaluate the collision integral numerically but never ask how a learned approximation distributes its errors across mechanisms.
Research Program
The data come from re-implementing the 2D resonant-manifold evaluator of Wu et al., together with the induced-diffusion triad selection criterion of Wu and Pan. The criterion is scale separation in vertical wavenumber and frequency, with the first wave much larger in scale than the other two. Both the evaluator and the criterion keep the published assumptions of horizontal isotropy and weak nonlinearity. We will generate a family of power-law action spectra by varying the vertical-wavenumber slope, the frequency slope and the energy level around GM76. For each spectrum we store the total collision integral and its decomposition into four classes: local interactions, parametric subharmonic instability, elastic scattering and induced diffusion. Before any learning, we will check that the re-implementation reproduces two published values. The first is the GM76 net downscale flux of about $1.5\times10^{-9}$ W kg$^{-1}$ (Wu et al.). The second is the blue-spectrum induced-diffusion flux of $(20.99\pm3.10)\times10^{-10}$ W kg$^{-1}$ (Wu and Pan).
The surrogate maps a discretized action spectrum $n(k_h, k_z, \omega)$ to the rate of change of the spectrum, $\partial n/\partial t$. Here $k_h$ is horizontal wavenumber, $k_z$ is vertical wavenumber and $\omega$ is frequency. The local branch sees the spectrum alone. The scale-separated branch also sees the action-spectrum gradients $\partial n/\partial k_z$ and $\partial n/\partial \omega$, and it is trained on the solver's induced-diffusion component. The remaining classes are lumped into the local branch's target. Two further steps reuse ideas from the debiasing literature. First, during training we record when each interaction class's error falls below a threshold, which gives a SEBRA-style learning-order measurement for each mechanism. Second, as a secondary ablation, we reweight training spectra by the failure of a deliberately shallow surrogate, in the manner of Learning from Failure. The goal is to test whether generic hard-example reweighting can substitute for physical structure.
Baselines
The naive reference point is the finescale parameterization, the standard formula oceanographers use to estimate dissipation from shear and strain. Wu et al. report that the true GM76 flux is about 1.5 times the finescale estimate. No reported result was found on Lacuna for the finescale formula on red or blue spectra, so we will compute it on our spectrum family ourselves.
The accuracy ceiling is the direct numerical evaluation of the collision integral by Wu et al. and Wu and Pan. It is the strongest existing method, and it also serves as ground truth. A surrogate "beats" it only on cost. The success condition is a large wall-clock speedup at a flux error well inside the reported $\pm3.10\times10^{-10}$ W kg$^{-1}$ uncertainty on the blue spectrum. No reported result was found on Lacuna for any neural surrogate of internal-wave collision integrals.
The strongest machine-learning competitor is a monolithic surrogate with random Fourier-feature inputs, trained on the total collision integral. Fourier features are the standard remedy for the spectral bias documented by Rahaman et al.. A plain monolithic network will also be run. For fairness, every surrogate receives the same training spectra, a matched parameter count, the same number of optimizer steps and the same hyperparameter search budget. The proposal counts as better if it cuts the error on induced-diffusion dissipation share substantially on held-out extreme slopes while matching total-flux error on GM76.
Two ablations isolate the key ingredients. The first is a split surrogate without the gradient inputs, which tests whether priming matters beyond the decomposition itself. The second is a local-only surrogate with no induced-diffusion branch. It should be nearly exact on GM76 and fail in proportion to the induced-diffusion share, up to the 25% reported by Wu and Pan.
Why not just…?
Why not just use Fourier features or a bigger network and train longer? This is the most serious threat. If spectral bias were the whole story, Fourier features would fix it without any physics. Our argument is that the difficulty of induced diffusion comes from scale separation, a coupling between distant parts of the spectrum, and not only from high frequency. A coordinate remapping does not obviously expose that coupling. The Fourier-feature monolithic baseline, run at matched budget and also at a several-times-larger budget, settles the question directly. If it closes the induced-diffusion gap, we will report that priming is unnecessary.
Why not just train per-class surrogates on the solver's decomposition, without priming? The solver already labels every mechanism, so supervision alone might suffice. The ablation that keeps the split but removes the gradient inputs measures exactly this. The physical motivation for priming is that induced-diffusion flux is controlled by the action-spectrum gradient and vanishes when the gradient vanishes (Wu et al.). A network that must rediscover that dependence from raw spectra should extrapolate worse to unseen slopes. If it does not, the contribution reduces to a decomposition-based surrogate, and we will say so.
Why not just reweight hard spectra with Learning from Failure and skip the physics? Generic reweighting needs no domain knowledge. However, SEBRA notes that this family of methods confuses intrinsically noisy examples with informative ones. In our setting, high-wavenumber regions where weak nonlinearity is questionable could absorb the extra weight instead of induced-diffusion regions. Running LfF reweighting on the monolithic surrogate tells us whether it targets the right mechanism. We check this by correlating the learned weights with the solver's induced-diffusion share.
Why not just run direct simulations of the Boussinesq equations? Direct simulation resolves every wave phase but cannot attribute flux to specific triad classes. Attributing flux to mechanisms is precisely what made the local-versus-induced-diffusion findings possible (Wu et al.). Direct simulation is also too costly to run inside the spectrum sweeps we target.
Proposed Experiments
The first decisive check is the learning-order measurement. If the class-wise convergence times of the monolithic surrogate do not place induced diffusion last, hypothesis two fails. The proposal would then rest on architecture alone.
The second check is extrapolation. We train on slopes near GM76 and test on the extreme red and blue spectra from Wu and Pan. Metrics are relative error in net downscale flux, error in induced-diffusion dissipation share, and band-wise transfer error in $k_z$ and $\omega$. As a sanity check, every surrogate must also reproduce the near-zero induced-diffusion flux on GM76.
The third check addresses the stationarity limitation that both Wu papers leave open. We will integrate the wave kinetic equation forward from perturbed non-equilibrium spectra, once using the solver and once using each surrogate, and compare spectral trajectories and the accumulated energy dissipated over short horizons. Errors that look small on single evaluations can compound when the surrogate is stepped repeatedly. A surrogate that drops induced diffusion should relax toward the wrong state, which a single-evaluation test cannot reveal.
We will also report wall-clock speedup against the 2D manifold evaluator on identical hardware. All experiments stay within the weak-nonlinearity regime in which the ground truth is valid. Spectra whose steepness violates that regime are excluded and reported as excluded, not scored.
References
- Energy cascade in the Garrett–Munk spectrum of internal gravity waves
- Induced diffusion of internal gravity waves: directionality and role in ocean mixing
- On the Spectral Bias of Neural Networks
- Fighting Fire with Fire: Avoiding DNN Shortcuts through Priming
- Towards Debiasing NLU Models from Unknown Biases
- SEBRA : Debiasing through Self-Guided Bias Ranking
- Learning from Failure: De-biasing Classifier from Biased Classifier