Insights

Data Meets Physics: The Next Wave in Oilfield Innovation

Xecta Team

May 9, 2024

We are in the age of AI. Advances in AI are happening at breakneck speed. Generative AI has captured people’s imaginations that only rival science fiction. But what led us here? What has happened over the last 5 to 10 years in terms of machine learning (ML) evolution?

Of course, the confluence of several factors related to data and computing contributed to AI’s acceleration. One of the key factors is the embedding of “structure” within the ML architectures. As we went deeper with the neural networks, we also encoded more structure within these algorithms (e.g. CNN, RNN, LSTM, Auto-encoder, GAN, Transformers) to improve accuracy and their expected behavior. However, most of these deep neural network methods are data hungry and require vast amounts of relevant data for training.

But how can we approach domain-specific computational problems when we do not have much data for training these data models and yet be reliable? In upstream applications, we often deal with large dimensional systems (reservoirs) that are sparsely sampled (by drilling a few wells), and we collect some data (say rates, pressures etc.). However, we may understand the governing behavior of underlying physical and chemical processes. Specifically, how can we then embed the structure of physics within these data models?

That brings us to hybrid models – also, referred to as Scientific Machine Learning (SciML). Hybrid models are physics-constrained data-driven models that are designed to work with lesser amounts of data. They balance simulation or data acquisition cost and accuracy to solve specific problems. Hybrid models are addressing similar needs in several disciplines spanning computational fluid dynamics, protein folding studies, climate modeling, aerospace, bioengineering, oil and gas, automotive engineering, process modeling etc.

Broadly speaking, hybrid models can be classified into 3 types:

1. Reduced order models

Start with a physics model and reduce its complexity, while preserving its important dynamics. For instance, reservoir simulation can be made to run faster with neural operators, modal decompositions etc.

2. Reduced physics

Build an approximate physics model that have self-learning features that are fit-for-purpose. For instance, flow network models such as Reservoir Graph Network (RGNet), Capacitance Resistance Model (CRM) etc. can be used for reservoir flood management applications.

3. Data physics

Starts with an ML model but constrain the solution space to honor physical laws. For instance, craft the model’s objective to account for unmodeled physics through a data-driven component.

Do hybrid models work? If so, where can they be applied in oil and gas?

Within reservoir and production domains, hybrid models are justified when pure physics models are either inadequate, slow, or expensive to characterize, build, and run to support the pace of decision-making. Fundamentally, we can reimagine several core engineering analyses using hybrid models. The following are a sample of scaled-up workflows with proven value cases.

Virtual BHP

We use multiphase flow correlations to estimate bottomhole pressure (BHP) from surface measurements (wellhead pressure and rates). However, selecting the best correlation and tuning them periodically is a cumbersome task if you must manage a large field. If some downhole pressure measurements are available, they could be used to train a classification algorithm to select the best correlation. Further, the unmodeled physics can be compensated by physics-informed machine learning (PIML) to improve the accuracy of the virtual BHP sensor.

Shutin Analysis

Pressure transient analysis (PTA) is the dominion of the reservoir engineer and involves interpretation. Hybrid models can be used to detect shutins; extract the reservoir signals during the build-up period to consistently extrapolate to obtain average reservoir pressure; perform simple buildup analysis to estimate near-wellbore reservoir parameters; and track well performance even when we do not have complete shutins.

Unconventional Forecasting

Predicting future flow rates reliably under various operating strategies for transient well flow is fundamental to establish benchmarks to track well performance. A reduced physics model tracks pressure depletion and reservoir performance that can be extrapolated to forecast multiphase flow rates. When well interference is detected, a multi-segment PI-based forecast further improves accuracy under modified reservoir conditions.

Artificial Lift Life Cycle Optimization

With rapidly changing well conditions in unconventional reservoirs, prompt timing and selection of the right artificial lift type amongst viable candidates can be evaluated under various operating strategies using hybrid models to recommend the best alternative. Any alternate scenario is compared against the base case i.e. the current lift type that is calibrated based on actual operating conditions. For the selected lift type, optimization algorithms can recommend optimal lift parameters (e.g. pump speed, gas lift injection rate etc.).

Waterflood Optimization

Flow network based reduced physics models (RGNet) are used to rapidly history match low-dimensional reservoir models from routine operational data (rates, pressures). This compact and fast model is used to estimate connected drainage volume, inter-well connectivity, forecasting, production control, and flood optimization to recommend injection and production rates. Any unmodeled physics is addressed with physics-informed machine learning (PIML) to improve forecasting accuracy.

Final thoughts

While hybrid models are still at their infancy in oil and gas adoption, the results point to a promising future. It is conceivable to imagine such models to advance in terms of accuracy and scope of applications in near future.