Abstract
In this paper, we propose to numerically assess the performance of standard Gaussian approximations to probe the posterior distribution that arises from Bayesian data assimilation in petroleum reservoirs. In particular, we assess the performance of (i) the linearization around the maximum a posteriori estimate (LMAP), (ii) the randomized maximum likelihood (RML), and (iii) standard ensemble Kalman filter-type (EnKF) methods. In order to fully resolve the posterior distribution, we implement a state-of-the-art Markov Chain Monte Carlo (MCMC) method that scales well with respect to the dimension of the parameter space. Our implementation of the MCMC method provides the gold standard against which to assess the aforementioned Gaussian approximations. We present numerical synthetic experiments where we quantify the capability of each ad hoc Gaussian approximation in reproducing the mean and the variance of the posterior distribution associated with a data assimilation problem. The main objective of our controlled experiments is to exhibit the substantial discrepancies of the approximation properties of standard Gaussian approximations.
Introduction
In petroleum engineering, "data assimilation"—often referred to as history matching—is the critical process of updating reservoir models with observed production data, such as flow rates and pressure. Given that subsurface characteristics like permeability cannot be directly observed, these models are characterized by significant uncertainty. The Bayesian framework provides a mathematically rigorous approach to this problem, yielding a posterior distribution that represents all reservoir states consistent with observed data.
However, the high dimensionality and extreme nonlinearity of reservoir models make exact Bayesian inference computationally prohibitive. Consequently, practitioners rely on "ad hoc" Gaussian approximations—mathematical shortcuts that assume uncertainty follows a Gaussian (bell-shaped) distribution. While computationally efficient, the accuracy of these approximations has historically lacked a rigorous benchmark. This study addresses this gap by comparing popular Gaussian methods against a high-fidelity Markov Chain Monte Carlo (MCMC) solution.
Methodology: Establishing the Gold Standard
To evaluate the approximations, the authors implemented a preconditioned Crank-Nicolson (pCN) MCMC method. Unlike standard MCMC algorithms, pCN is designed to scale efficiently with high-dimensional parameter spaces. By running millions of simulations (approximately $5.5 \times 10^7$), the authors were able to fully resolve the posterior distribution, providing a "gold standard" against which other methods could be measured.
The study compared this benchmark against three primary classes of Gaussian approximations:
- Linearization around the Maximum A Posteriori (LMAP): This method finds the single most likely model and assumes the uncertainty is linear in its vicinity.
- Randomized Maximum Likelihood (RML): This approach generates an ensemble by solving multiple randomized optimization problems.
- Ensemble Kalman Filter (EnKF) and Ensemble Square-Root Filter (EnSRF): These are sequential methods that update an ensemble of models as data becomes available. The study also evaluated versions using distance-based localization to mitigate sampling errors.
Experimental Design
The performance of these methods was tested using synthetic 2D reservoir models to ensure ground truth was known:
- Single-phase Darcy flow: A relatively less nonlinear model involving only oil.
- Two-phase oil-water reservoir: A highly nonlinear model representing a more realistic and complex challenge.
- Well Configurations: Scenarios varied from sparse data (5-spot pattern) to dense data (25 wells) to observe how data density influenced the approximations.
Key Insights and Comparative Analysis
The numerical experiments revealed several critical discrepancies between the approximations and the true posterior:
The Impact of Nonlinearity
The physics of fluid flow in porous media is inherently nonlinear. The study confirmed that even when starting with a Gaussian prior, the resulting posterior distribution is decidedly non-Gaussian. This fundamental mismatch is the primary reason Gaussian approximations struggle to capture the true state of the reservoir.
The Robustness of LMAP
One of the most surprising findings was the performance of LMAP. Despite being the simplest approximation, it was often the most robust at reproducing both the mean and the variance of the posterior, particularly in complex two-phase models. In the two-phase scenario with few wells, LMAP achieved a relative error in the mean of 0.284, significantly outperforming more complex ensemble methods.
RML and the Variance Problem
While RML proved excellent at identifying the "mean" or average state of the reservoir—achieving a relative error of 0.154 in single-phase models—it struggled with uncertainty quantification. In highly nonlinear cases, RML tended to significantly overestimate the variance (uncertainty), potentially leading to incorrect engineering decisions regarding risk.
Limitations of EnKF
The Ensemble Kalman Filter (EnKF) performed poorly with standard ensemble sizes due to "spurious correlations." Even with the introduction of localization and square-root filters, EnKF failed to converge to the true posterior, even as ensemble sizes were increased to thousands of members. In the two-phase model, EnKF's error in the mean was a staggering 1.159, though localization reduced this to 0.600.
Conclusion
This research highlights the significant trade-offs inherent in reservoir data assimilation. While Gaussian approximations like LMAP and EnKF offer massive computational savings—requiring only dozens or hundreds of runs compared to the millions required for MCMC—they do so at a cost to accuracy.
The study suggests that LMAP is a surprisingly strong contender for uncertainty quantification in nonlinear systems. However, the authors note that their study focused on unimodal distributions. Future work will extend these investigations into 3D models and explore more sophisticated localization techniques to bridge the gap between computational efficiency and Bayesian rigor.
Summary updated