Mapping of hydraulic transmissivity field from inversion of tracer test data using convolutional neural networks. CNN-2T
Résumé
This paper introduces a new concept for mapping hydraulic transmissivity from temporal concentration data collected in multiple tracer tests. Based on convolutional neural network, the principle uses an encoder-decoder architecture with multiple neural layers to establish a relationship between the concentration data and the transmissivity field. This relationship is established in two phases with two networks. The first network is designated and trained to reconstruct a transmissivity field using data from a single tracer test. To improve the reconstruction quality, the second network then performs a joint interpretation for multiple tracer tests, which reprocesses all the transmissivity resulted from the first network for each individual tracer test. Both networks are trained by synthetic data, where the transmissivity models are generated with a Gaussian variogram and its properties are considered as prior information on the aquifer heterogeneity. Tracer tests are derived numerically by solving the forward problem to obtain the corresponding concentration data that feed the training. The trained networks accurately map the transmissivity fields, of which the accuracy relies on the volume and nature of the heterogeneities of training models, as well as the number of piezometers used to monitor the concentration changes. Reconstruction quality, on the other hand, is less influenced by data noise. Effective training requires a large dataset, but the time required for dataset generation is only on the order of the Gauss-Newton algorithm in a conventional inversion, while the trained network performs inference instantly.