Machine learning applied to pumping test analysis

Aquifer test analysis is complex, and in many regards, the interpretation resembles an art more than a science. Under the best circumstances, aquifer test analysis is still plagued by ambiguity and uncertainty, compounded by the general lack of information on the subsurface. An approach which has seen widespread adoption in other fields that need to classify time series data is machine learning. A Python script that generates numerical groundwater flow models by interfacing directly with the modelling software produces training data for deep learning. Production yielded 3,220 models of aquifer tests with varying hydrogeological conditions, including fracture, no-flow and recharge boundary geometries. Post-processing exports the model results, and the Bourdet derivative is plotted and labelled for image classification. The image classifier is constructed as a simple three-layer convolutional neural network, with ReLU as the activation function and stochastic gradient descent as the optimizer. The dataset provided sufficient examples for the model to obtain over 99% accuracy in identifying the complexities present inside the numerical model. The classification of groundproofing data illustrates the model’s effectiveness while supporting synthetically prepared data using modern groundwater modelling software.