Data Driven Air Entrainment Velocity Parameterization by Breaking Waves

Wave breaking injects turbulence and bubbles into the upper ocean, modulating air-sea exchange of momentum, heat, gases, and sea-spray aerosols. These fluxes depend nonlinearly on sea state but remain poorly represented in coupled atmosphere-wave-ocean models, where air-entrainment velocity is often parameterized using wind speed or significant wave height alone. We develop a global machine-learning parameterization of Va trained on a 43-year WAVEWATCH III simulation that resolves the breaker-front distribution and associated energetics. A multilayer perceptron with seven physically motivated predictors (wind speed, wave height, wave age, steepness, direction, and depth) reproduces spectral-reference Va with high skill. The model reduces longstanding biases in bulk formulas, notably overestimation in swell-dominated low latitudes and underestimation in storm tracks. Applied globally, it improves bubble-mediated CO2 transfer velocity and sea-salt aerosol emission, reducing errors by an order of magnitude. Validation against independent HiWinGS observations supports robust deep-water performance.

💡 Research Summary

This paper addresses the longstanding challenge of representing wave‑breaking‑driven air entrainment in coupled atmosphere‑wave‑ocean models. Air‑entrainment velocity (Vₐ) controls the injection of turbulence and bubbles that strongly enhance momentum, heat, gas, and sea‑spray exchanges at the ocean surface. Traditional parameterizations rely solely on wind speed (U₁₀) or, at best, a semi‑bulk combination of wind speed and significant wave height (Hₛ). Such approaches ignore the rich sea‑state dependence (wave age, steepness, direction, depth) and consequently produce systematic biases: overestimation in swell‑dominated tropics and underestimation in mid‑latitude storm tracks.

To overcome these limitations, the authors generate a reference dataset using a 43‑year global hindcast with WAVEWATCH III (WW3 v7.14) at 0.5° resolution. The model implements the breaking‑crest distribution Λ(c) via the ST4 source term, allowing a physically based calculation of Vₐ from the integral Vₐ = ˜B ∫c³⁄² S(k)³⁄² Λ(c) dc (˜B = 0.1). The hindcast provides 3‑hourly fields of Vₐ, Hₛ, friction velocity, wave age (cₚ/U₁₀), steepness (kₚHₛ/2), wind direction, and water depth for the period 1980‑2022.

A multilayer perceptron (MLP) is trained to emulate this reference Vₐ using seven readily available predictors: Hₛ, U₁₀, cosine and sine of wind direction, wave age, steepness, and depth. The network architecture consists of four fully‑connected hidden layers (512 neurons each) with GELU activations and 10 % dropout, followed by a linear output. Training employs the Adam optimizer (lr = 10⁻³, weight decay = 10⁻⁴) in mixed‑precision, distributed across GPUs. Data are split into training (1980‑2014, 80 %), validation (2015‑2018, 10 %), and testing (2019‑2022, 10 %). After 12 epochs, the model achieves an RMSE of 0.11 cm h⁻¹, normalized RMSE of 0.08, and a correlation coefficient of 0.999 against the withheld WW3 reference, with a negligible bias (‑0.018 cm h⁻¹). By contrast, a wind‑only power‑law (Vₐ = a₂(U₁₀‑c₂)ᵇ) and a semi‑empirical friction‑height formulation (Vₐ = a₁Cₚ(u*/√gHₛ)ᵇ) yield RMSEs of 1.51 and 1.14 cm h⁻¹, biases of –0.279 and –0.081 cm h⁻¹, and lower correlations (0.92 and 0.87).

The authors further validate the MLP against independent observations from the HiWinGS field campaign in the North Atlantic. Using measured directional spectra, laser altimetry, and a local WW3 hindcast, they reconstruct Λ(c) and compute observed Vₐ. The MLP reproduces these observations with R = 0.76, RMSE = 64.2 cm h⁻¹, and a positive bias of 53.7 cm h⁻¹. Skill is highest under strong wind and steep‑wave conditions; the model tends to overestimate Vₐ in low‑wind, weak‑breaking regimes, reflecting the limits of the equilibrium‑based Λ(c) reconstruction and detection thresholds for micro‑breaking.

Applying the ML‑derived Vₐ globally (2019‑2022) demonstrates substantial improvements over bulk schemes. Zonal‑mean comparisons show that the ML model captures the pronounced latitudinal variability of the reference (high values in Southern Ocean and North Atlantic storm tracks, low values in subtropical gyres) with residuals < 0.05 cm h⁻¹. Bulk formulations underestimate Vₐ in mid‑latitude storm tracks and overestimate it in tropical swell zones. The improved Vₐ propagates into downstream diagnostics: bubble‑mediated CO₂ transfer velocity and sea‑salt aerosol emission estimates are corrected by roughly an order of magnitude, indicating that climate‑scale models could achieve far more realistic air‑sea fluxes without incurring the computational cost of full Λ(c) diagnostics.

The study emphasizes portability: all seven predictors are standard outputs of wave models, reanalyses, or satellite‑derived products, enabling straightforward integration of the ML parameterization into existing modeling frameworks without retraining. Limitations are acknowledged: (1) reliance on WW3 physics means the surrogate inherits any systematic errors of the source model; (2) performance degrades in non‑equilibrium, low‑wind conditions; (3) the model has not been tested under future climate scenarios or in polar regions with extensive sea‑ice cover. Future work should explore hybrid approaches that blend the ML surrogate with physics‑based corrections for low‑energy regimes and extend validation to high‑resolution field campaigns and climate projections.

In summary, the paper delivers a physically interpretable, high‑skill, computationally cheap surrogate for wave‑breaking‑driven air‑entrainment velocity. By leveraging a long‑term global wave hindcast and a modest neural network, the authors achieve a parameterization that markedly reduces biases of traditional bulk formulas, improves estimates of gas exchange and aerosol production, and offers a ready‑to‑use tool for the next generation of coupled Earth system models.