Subsurface flow data assimilation gets adaptive-tempered sequential Monte Carlo: posteriors match reference MCMC on four synthetic truths, with 2.5–7.0× single-GPU speedups
Synopsis
Geological uncertainty in subsurface flow systems is often hierarchical: a set of geological metaparameters controls overall statistics while cell-by-cell property fields vary, and both must be estimated from observations. An adaptive-tempered sequential Monte Carlo method estimates metaparameters and Gaussian latent variables jointly within one likelihood framework, with surrogate models supplying all forward evaluations. Across four synthetic true models, its posterior metaparameter distributions and flow predictions agree closely with reference MCMC results, while single-GPU data assimilation is 2.5–7.0 times faster and needs fewer function evaluations.
Interpretation
History matching for subsurface flow can now estimate geological metaparameters and cell-by-cell realizations together, instead of treating the metaparameters as fixed settings. Ensemble Kalman methods have typically generated prior realizations under fixed metaparameters, and their Kalman updates rely on finite ensemble covariance estimates and statistical linearization, so they are not guaranteed to produce exact samples from the target posterior for nonlinear, non-Gaussian problems. In two 3D systems with hierarchical uncertainty, a faulted carbon storage system and a channelized oil-water system, posterior metaparameter distributions and flow predictions agree closely with reference MCMC results for the four synthetic true models.
The method writes each geological realization as a combination of metaparameters and Gaussian latent variables that are independent under the joint prior. The noncentered geological parameterization lets metaparameters scale and shift properties without changing the spatial correlation structure, so the latent variables can follow a standard Gaussian prior. The faulted system represents target-aquifer heterogeneity with principal component analysis, and the channelized system represents channel geometry and facies realizations with a 3D latent diffusion model; assimilation is carried out in each 3D system.
A population of particles is transported from the joint prior to the hierarchical posterior, with only part of the likelihood introduced at each step. Adaptive tempering selects temperature increments from the effective sample size of the importance weights to limit particle weight degeneracy, which can arise when prior samples are weighted directly by the full likelihood. Each iteration applies importance weighting, systematic resampling and short-chain blockwise MCMC mutation, with preconditioned Crank–Nicolson proposals for latent variables and Gaussian random-walk proposals for metaparameters, and all function evaluations come from deep-learning flow surrogates.
Perspective
The demonstrated scope is two 3D synthetic systems: a carbon storage system with three stacked aquifers and two faults, and a channelized oil-water system with three facies; observations are generated by high-fidelity simulation plus Gaussian noise, not field data. The authors propose future directions including active learning to improve surrogate accuracy, evaluation of multistage assimilation, use of updated posteriors for sequential decision-making, and treatment of model discrepancy.
Some metaparameters are only weakly constrained by the observations, with posteriors that barely shrink relative to their priors, and small local peaks in the kernel density estimates are sensitive to bandwidth, which the authors advise interpreting cautiously. Under the stricter convergence criterion, several MCMC calculations reached the 50-hour time limit without satisfying it, so some comparisons rest on the less stringent convergence point.