What are the benefits and limitations of using proxy models and machine learning in history match?

Proxy models can easily represent non linear behaviour and interactions through Gaussian processes. They have shown their benefit on hundreds of projects since 2001. I agree about extrapolation, like any model, but this can be corrected by suitable choice of sequential simulations. The main thing which my paper addresses, SPE 173301, is that uncertainty quantification needs to be done using Hamiltonian Markov Chain Monte Carlo processes, and the uncertainty quatification needs to be represented by actual simulation runs, not by the proxy model. Take a look.

Proxy models serve as efficient approximations of complex reservoir simulations, enabling rapid evaluations through mathematical functions like polynomials or Gaussian processes. While they significantly reduce computational demands and facilitate tasks such as sensitivity analysis and optimization, their limitations include potential inaccuracies in capturing nonlinearities and interactions inherent in the reservoir system. Regular validation and updates are essential to maintain their reliability, particularly when extrapolating beyond the training data range. Thus, while proxy models are invaluable tools, they must be employed judiciously to ensure accurate decision-making in reservoir management.

Machine learning, a branch of artificial intelligence, enhances history matching in reservoir modeling by automating tasks like proxy model generation and simulation run classification. It can infer model parameters directly from production data using inverse modeling, streamlining workflows and uncovering insights. However, its effectiveness relies on having substantial, reliable data and a solid understanding of the reservoir's physical and geological context to avoid overfitting and bias. Thus, careful selection and validation of machine learning methods are crucial for achieving realistic results in reservoir management.

For more than a handful of parameters, Hamiltonian methods are essential, and these need proxy models in order to be done in a practical time frame. Anything else is a possibilistic, rather than probabilistic, approach.

The history matching process is a crucial step in reservoir modeling, as it helps in accurately estimating the future performance of the field and developing optimal production plans. One of the major challenges is balancing the time required to run the model with achieving optimal history matching results. Consequently, uncertainty analysis becomes essential in reservoir simulation to enhance the accuracy of history matching. Proxy models and machine learning provide significant benefits in this context, including reduced computational cost, faster iterations, efficient handling of large datasets, integration of diverse data sources, and automated workflows

History match quality and uncertainty analysis are critical for assessing the accuracy of reservoir simulations and the reliability of model predictions. Utilizing proxy models allows for efficient exploration of parameter spaces to identify optimal history match solutions and quantify output uncertainties. Machine learning further enhances this process by evaluating history match solutions against various criteria and dynamically updating model parameters as new data emerges. Together, these approaches improve decision-making and risk management in reservoir management by providing more robust insights into model performance and uncertainties.

This is a great article about the usefulness of proxy models and machine learning in automating and speeding up the history matching process for 3D numerical reservoir models. Simulation engineers need a good understanding of the physics of reservoir fluid flow coupled with an understanding of the quality of the available surface production and pressure data. When the engineer understands reservoir and well-bore flow hydraulics, they will be able to critically weight all the data that is fed to the numerical models. They would also be able to quickly realize that certain combinations of reservoir parameters are not physically possible even if they provide a numerical ′match’ and thus design more realistic what-if investigative scenarios.

In addition to enhancing history match quality and uncertainty analysis, it is essential to consider the integration of multidisciplinary approaches in reservoir modeling. Collaborating with geologists, engineers, and data scientists can provide a more comprehensive understanding of reservoir behavior and improve model accuracy. Furthermore, leveraging advanced data analytics and visualization techniques can facilitate better communication of insights and foster informed decision-making. Ultimately, a holistic approach that combines technical expertise with innovative methodologies will lead to more effective reservoir management strategies.