Developing a reliable deep learning model for new unconventional reservoirs, is often constrained by the limited number of wells available. Transfer learning is a useful approach to alleviate data needs in training a neural network. This involves storing knowledge gained while solving one problem (source data) and applying it to a different but related problem (target data). However, the transfer of incorrect knowledge can impede the performance of the model trained for the new field. Furthermore, the black-box nature of such networks makes it difficult to interpret the contribution of the multiple source models used for knowledge transfer. Hence, ranking their viability for transfer can help with avoiding negative transfer and improving the predictive performance in the target dataset. After identifying the appropriate network, we propose a neural network structure to combine multiple source models into one network to be transferred to a target dataset. Each source model is individually trained using their corresponding data and their related weights are fixed during retraining for the new target dataset. This approach allows for combining the outputs of different source models in describing the target dataset. We first illustrate the impact of negative transfer when an incorrect source model is used for transfer learning. We then apply the proposed approach to a synthetic dataset to show how different sources models are ranked by a classifier. The ranking is obtained by comparing the average probabilities extracted from the SoftMax activation on all the test cases for the target dataset. Once the appropriate models are selected, we study the sensitivity to the number of data points available for retraining to highlight the importance of transfer learning with the aggregated framework.
Unconventional reservoirs in the United States are typically tight oil formations that have extremely low permeabilities. The drilling and completion techniques used for conventional high permeability reservoirs are not practical for such formations. However, utilizing key technologies, like hydraulic fracturing and horizontal drilling, has allowed for the economical production of hydrocarbons from this low permeability shale (King, 2010).
