Abstract
Most of the bulk algorithms used to calculate turbulent air–sea fluxes of momentum and heat are iterative algorithms whose convergence is slow and not always achieved. To avoid these drawbacks that are critical when large datasets must be processed, a statistical model of bulk air–sea fluxes based on artificial neural networks was developed. It was found that classical bulk algorithms were slower than the statistical model, by a factor of 1.75–7 depending on the bulk algorithm selected for the comparison. A set of 12 global analyses of an operational meteorological model as well as in situ data corresponding to equatorial and midlatitude conditions were used to assess the accuracy of the proposed model. The wind stress, latent, and sensible heat fluxes calculated with neural networks have acceptable biases with respect to bulk fluxes, between 0.4% and 1% depending on the flux magnitudes. Moreover, the rms deviation between bulk fluxes and neural network flux estimates is only 0.003 N m−2 for the momentum flux, 0.5 W m−2 for the sensible heat flux, and 1.8 W m−2 for the latent heat flux, at global scale, which is small compared with the natural variability of these quantities or the expected error.
Corresponding author address: Denis Bourras, CETP-IPSL-CNRS 10-12, Avenue de l’Europe, 78140 Vélizy-Villacoublay, France. Email: denis.bourras@cetp.ipsl.fr
