Abstract
Accurately quantifying air-sea fluxes is important for understanding air-seainteractions and improving coupled weather and climate systems. This studyintroduces a probabilistic framework to represent the highly variable nature ofair-sea fluxes, which is missing in deterministic bulk algorithms. AssumingGaussian distributions conditioned on the input variables, we use artificialneural networks and eddy-covariance measurement data to estimate the mean andvariance by minimizing negative log-likelihood loss. The trained neuralnetworks provide alternative mean flux estimates to existing bulk algorithms,and quantify the uncertainty around the mean estimates. Stochasticparameterization of air-sea turbulent fluxes can be constructed by samplingfrom the predicted distributions. Tests in a single-column forced upper-oceanmodel suggest that changes in flux algorithms influence sea surface temperatureand mixed layer depth seasonally. The ensemble spread in stochastic runs ismost pronounced during spring restratification.