Function-Space Decoupled Diffusion for Forward and Inverse Modeling in Carbon Capture and Storage

2026-02-12 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Geophysical Modeling & Simulation · Depth: Expert, quick

Summary

Fun-DDPS is a novel generative framework designed for accurate subsurface flow characterization in Carbon Capture and Storage (CCS) applications, addressing challenges posed by ill-posed inverse problems with sparse observations. It integrates function-space diffusion models with differentiable neural operator surrogates for both forward and inverse modeling. The system learns a prior distribution of geological parameters using a single-channel diffusion model and employs a Local Neural Operator (LNO) surrogate for physics-consistent guidance. This decoupled architecture allows robust recovery of missing parameter space information and efficient gradient-based data assimilation. On synthetic CCS datasets, Fun-DDPS achieved an 11x improvement in forward modeling, reducing relative error to 7.7% with only 25% observations, compared to 86.9% for standard surrogates. It also demonstrated rigorous validation against Rejection Sampling posteriors for inverse solving, achieving a Jensen-Shannon divergence less than 0.06 and producing physically consistent realizations with 4x improved sample efficiency.

Key takeaway

For research scientists developing subsurface flow models in Carbon Capture and Storage, Fun-DDPS offers a significant advancement in handling data sparsity and inverse problem challenges. You should consider integrating this function-space decoupled diffusion approach to achieve higher accuracy in forward modeling and more efficient, physically consistent inverse solutions, particularly when dealing with limited observational data. This method can drastically reduce error rates and improve sample efficiency compared to traditional surrogates and rejection sampling.

Key insights

Fun-DDPS combines diffusion models and neural operators for robust, efficient, and physics-consistent CCS modeling.

Principles

• Decoupling parameter and dynamics fields improves robustness.
• Physics-consistent guidance enhances generative models.

Method

Fun-DDPS learns a geomodel prior via diffusion, then uses an LNO surrogate for physics-consistent, cross-field conditioning to guide data assimilation in CCS modeling.

In practice

• Apply Fun-DDPS for CCS with sparse observation data.
• Use LNO surrogates for physics-consistent guidance.

Topics

• Function-Space Diffusion Models
• Neural Operators
• Carbon Capture and Storage
• Inverse Problems
• Forward Modeling

Best for: Research Scientist, AI Researcher, AI Scientist

Related on AIssential

• Latent Diffusion Posterior Sampling with Surrogate Likelihood Guidance for PDE Inverse Problems — L-DPS uses latent diffusion and a neural surrogate for efficient, accurate Bayesian inversion of PDEs.
• Task Aware Modulation Using Representation Learning for Upsaling of Terrestrial Carbon Fluxes — Integrating physical constraints with adaptive representation learning improves global carbon flux estimation accuracy.
• Author Correction: Mask-prior-guided denoising diffusion improves inverse protein folding — A published correction clarifies MapDiff's amino acid prediction accuracy in inverse protein folding.
• FVD: Inference-Time Alignment of Diffusion Models via Fleming-Viot Resampling — FVD uses a Fleming-Viot birth-death mechanism to prevent diversity collapse in diffusion model inference, improving sample quality and efficiency.
• Sampling sea state using a diffusion model — Diffusion models offer a computationally efficient path to probabilistic global sea state estimation.
• Colored Noise Diffusion Sampling — Diffusion model SDE inference can be optimized by dynamically allocating noise energy based on frequency resolution.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.