Data science and machine learning have received significant attention in recent years, primarily due to unprecedented growth in data acquisition and recent advances in computing technologies, together with the reported success stories about their deployment and applications from industry, academia, and research communities. In traditional science and engineering disciplines, physical laws and principles are the main pillars for deriving predictive models, whereas data is often used to tune model parameters (e.g., material properties of the system). Unlike physics-based prediction models that establish causal relations among the variables in a given physical system, data-driven models are derived from identifying and exploiting statistical patterns and relations in data. Data-driven models have their origin in applications where (physical) models either do not exist or are too complex to build. However, the success of data science in recent years has inspired many scientists to apply the underlying methods to the analysis and prediction in physical systems.
At the SEES lab, we integrate machine learning and predictive analytics tools with fluid flow, geological, and engineering insights to develop efficient prediction models for a number of subsurface energy and environmental applications. Examples of our current projects include deep convolutional neural networks for dimension reduction and data conditioning, novel neural network architectures for characterizing subsurface flow connectivity, direct integration of dynamic flow data under uncertain geologic scenarios in low-dimensional latent spaces, and dynamic neural networks for fault detection and model-predictive control in geo-energy systems. Our current projects are sponsored by the Department of Energy, Energy Industry, and Research Foundations.
