Data assimilation for subsurface flow using latent diffusion model parameterization: performance of ensemble-Kalman and Monte Carlo techniques

Guido Di Federico, Wenchao Teng, Louis J. Durlofsky|June 9, 2026arXiv

Key Takeaway

When using neural network-based model compression for inverse problems, traditional ensemble Kalman methods can fail due to nonlinearity—rigorous Monte Carlo sampling with surrogate models offers a more reliable alternative.

Summary

This paper tackles a key challenge in subsurface modeling: updating geological parameters to match well observations while keeping models realistic. The authors use latent diffusion models to compress high-dimensional geological data into a smaller space, then compare different update methods (ensemble Kalman, MCMC, SMC).

training efficiency evaluation

Key Terms

data-assimilation latent-diffusion-models ensemble-kalman-filter markov-chain-monte-carlo surrogate-model