École Nationale Supérieure de
Techniques Avancées

Résumé

Studies of atmospheric dispersion of pollutants on a local scale are usually done with computational fluid dynamics (CFD) simulations. However, CFD computations are expensive, so can only be performed on a limited number of cases. Machine Learning (ML) approaches offer a promising alternative for faster results, including results of cases not present in the initial CFD-generated database. Here, a multi-layer perceptron (MLP) is trained on CFD simulation results with varying wind directions. The impact of the preprocessing steps is studied. Then, a sub-sampling method of spatial points is presented. The method allows to greatly reduce training time with little to no impact on the accuracy of the training and validation by selecting only a small number of spatial points from what is available in the CFD mesh. Finally, the trained MLP is compared to a simple interpolation-based baseline. It is shown that the model is more accurate and has better generalization capabilities

Versions disponibles de ce document

• Neural surrogates for atmospheric dispersion in built areas (deposited 14 nov. 2023 14:51) [Actuellement Affiché]
  • Neural surrogates for atmospheric dispersion in built areas (deposited 15 nov. 2023 15:57)