Multi-tasking
In this work, we present a novel method to simultaneously map transmissivity and storativity in a heterogenous aquifer using a multitask convolutional neural network. This coupled inversion algorithm translates transient hydraulic head data from pumping tests into two independent tomography of transmissivity and storativity in a two-dimensional problem.
Multi-task architecture in X-shape composing two encoder-decoder network, 95 layers with 0.6 million learnable parameters, with multiple linking and bridging mechanisms
Based on the SegNet architecture, the multitask neural network provides an effective solution with minimal weights and biases, works as an end-to-end operator that directly approximates the inverse function. Multiple sharing mechanisms enable the multitasking approach to outperform single-task models by 10% accuracy.
Representative examples from simple to complex. The proposed network provides accurate reconstructions for both fields, with accuracy depending on the complexity of target fields
Application to synthetic experiments shows that the quality of inverted maps relies on the complexity level of heterogeneity in the hydraulic fields. Mapping accuracy also depends on the data coverage, but being resistant to data noise due to the featuring mechanism in the convolutional network adopted. Comparison with conventional inversion method reveals that the deep learning approach provides better generalization in the inversion while performing each inference on the order of milliseconds without further calibration from users.
Links:
