INVERSION OF A SEMI-PHYSICAL DISPERSION MODEL

This study proposes to examine the performances of an inverse dynamic model resulting from the fusion of deterministic modeling and statistical learning. An inverse semi-physical or gray-box model is then carried out using a recurrent neural network (NN). The suggested model concerns in particular a pollutant dispersion phenomenon governed by a partial differential equation (PDE), on a basic mesh. This technique leads to the realization of a neural network inverse problem solver (NNIPS). The network is structured by the discrete reverse-time state form of the direct model. The performances are numerically analyzed in terms of generalization, regularization and training effort.