Cross-Domain Feature learning and data augmentation for few-shot proxy development in oil industry

Abstract

In reservoir engineering, numerical simulators are crucial for analyzing risks and uncertainties. The decision-making plan is complex due to numerous uncertain variables, which lead to high dimensional spaces, such as porosity and permeability 3-dimensional inputs, and a substantial computational footprint. The scientific literature has directed efforts toward proxy model construction that operates as a faster estimation function. Yet, many samples are needed to develop a reliable solution, increasing the proxy construction time. Our investigation introduces a novel Few-Shot Proxy method, which opens the possibility of dealing with high-dimensional inputs, decreasing simulator dependence without incurring a high computational cost. Our methodology utilized two initial sample sets comprising only 30 and 40 examples each. The data augmentation method ensures the requisite sample diversity essential for robust training. The study leverages the simulator’s reservoir uncertainties collection to supplement the variety and representativeness of the training instances. The Few-Shot Proxy method builds upon synthetic data modeling in a Cross-Domain Feature scheme. New training data are assembled based on a meticulous selection of features, underpinned by clustering and metric learning. For time series construction, we exploited the notion of fuzzy logic and prototypical networks to cover the sparse distribution. The proxy method generates the cumulative fluid production curve from geostatistical realizations and provides the risk curve considering an uncertainty distribution. The presented strategy decreased the simulator’s processing time by 87%, evaluating the training, validation, testing, and risk analysis steps. In the risk analysis, the symmetric mean absolute percentage error was less than 2% in all cases.