Do determinístico ao aprendizado de máquina: uma nova abordagem para estimar a turbulência atmosférica

Abstract

The representation of atmospheric turbulence in the Planetary Boundary Layer is one of the main sources of uncertainty in numerical weather and climate models. In this work, the use of artificial neural networks of the Multi-Layer Perceptron type is proposed as surrogate models to emulate the Holtslag–Boville turbulence parameterization scheme in the Brazilian one-dimensional model BAM-1D. The neural networks were trained using data generated by BAM-1D simulations forced by observations from the GoAmazon 2014/15 campaign, using normalized atmospheric variables as inputs and turbulent coefficients (kvm, kvh) and counter-gradient terms (cgs, cgh), as well as the Planetary Boundary Layer height, as outputs. The implementation of the networks was carried out directly in Fortran, allowing their integration as drop-in replacements for the original physical scheme. The results indicate that the neural network reproduces with high fidelity the vertical structure and temporal variability of the turbulent coefficients, presenting high coefficients of determination for kvm and kvh, as well as low statistical errors (RMSE, MAE, and bias). The analysis of temperature and wind profiles demonstrates that replacing the HB scheme does not compromise the thermodynamic and dynamic consistency of the model, with differences mainly concentrated near the surface and in highly convective regimes. Additionally, the evaluation of precipitation as an integrated metric indicates that the neural network preserves the occurrence and temporal phase of events, although differences in the intensity of convective peaks are observed, reflecting the nonlinear nature of the process. From a computational perspective, the machine learning-based approach shows potential for cost reduction while maintaining physical consistency. The results confirm the feasibility of using neural networks as surrogate models in atmospheric parameterizations, opening perspectives for the development of more efficient and accurate hybrid models.