A neural network emulator for the planetary boundary layer height

Abstract This study develops a data-driven method to emulate the diagnosis of the Planetary Boundary Layer height (PBLH) using a Multi-Layer Perceptron neural network (MLP-NN). The model is trained with simulations from BAM-1D, the one-dimensional version of the Brazilian Atmospheric Model, in which the Holtslag–Boville (HB) parameterization provides the reference PBLH diagnosis. Once trained, the neural network is able to reproduce this diagnosis without relying on observational data, using only variables available within the model itself. The MLP-NN accurately emulates the HB scheme, achieving a correlation of 0.99 during independent validation and maintaining very small generalization errors. The approach also demonstrates stable performance across contrasting diurnal cycles. Although the HB method remains faster in the tested configuration, the neural network offers a fixed-cost inference step that is favorable for future applications involving more complex scans or accelerator-based computation. Overall, the proposed MLP-NN provides a faithful and portable surrogate for the PBLH diagnosis in BAM-1D, establishing a foundation for future studies on efficiency-accuracy trade-offs and the integration of machine-learning emulators in atmospheric modeling systems.