Identifying and Ranking Multiple Source Models for Transfer Learning in Unconventional Reservoirs.

Abstract

When a limited number of wells are drilled at the early stages of developing unconventional fields, the available data is insufficient for developing data-driven models. To compensate for the lack of data in new fields, transfer learning may be adopted by using a previously learned model/knowledge from similar fields (source data) to build a predictive model for the new field. To be effective, transfer learning requires the source and target fields to have similarities and to ensure relevant information/knowledge is transferred. The transfer of irrelevant knowledge may impede the training process and lead to a negative knowledge transfer. When multiple source data are available, it is important to identify each source data's relevance and potential contribution to the target data. We introduce a framework to rank different source datasets and determine their capability for transfer learning. The methodology relies on using knowledge learned from datasets with similar features to the target dataset. This methodology helps circumvent the data needs for training while ascertaining that the right knowledge is transferred when developing new fields. Additionally, the framework allows for combining relevant features from multiple source models (with similar ranks). It allows for transferring the knowledge learned from mature fields to improve the performance of deep learning proxy models for new fields with similar features.