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

This paper introduces Fun-DDPS, a novel generative framework that decouples function-space diffusion priors from differentiable neural operator surrogates to achieve high-accuracy, physically consistent forward and inverse modeling for Carbon Capture and Storage, significantly outperforming standard surrogates in data-sparse scenarios and matching exact sampling posteriors with superior efficiency.

The Gist

Imagine you are trying to figure out what a hidden underground world looks like, but you can only peek through tiny, scattered holes in the ground. This is the challenge of Carbon Capture and Storage (CCS). Scientists need to know exactly where to store massive amounts of CO2 underground to keep it safe, but the ground is a complex, messy mix of rocks and fluids. They have very few measurements (sparse data) and need to predict how the gas will move (forward modeling) or guess the rock structure based on where the gas is found (inverse modeling).

The paper introduces a new AI tool called Fun-DDPS to solve this. Here is how it works, explained through simple analogies.

The Problem: The "Guessing Game" with Bad Clues

Imagine you are trying to reconstruct a shattered vase, but you only have two tiny shards and a blurry photo of the room.

• Old Methods (The "Fill-in-the-Blanks" Approach): Traditional AI tries to guess the whole vase by just filling in the empty spaces with "average" guesses. If the vase has a unique, jagged shape, the AI just smoothes it out, losing all the important details. It's like trying to draw a detailed map of a city using only a few street names; you end up with a blurry mess.
• The "Joint" Approach: Some newer AI models try to learn the shape of the vase and the room at the same time. But if they don't have enough examples, they get confused. They might start inventing weird, fizzy textures (artifacts) that look like static on an old TV, because they are trying to force a connection between the vase and the room that doesn't actually exist physically.

The Solution: Fun-DDPS (The "Specialized Team")

The authors propose Fun-DDPS, which is like hiring a specialized team of two experts instead of one generalist who tries to do everything. They "decouple" (separate) the tasks:

1. The Geologist (The Diffusion Prior)

• Role: This expert is an artist who knows exactly what underground rock formations look like. They have studied thousands of real geological maps.
• How they work: They don't care about the physics of gas flow yet. Their only job is to generate a realistic-looking map of the rocks. If you give them a blank canvas, they can paint a plausible rock formation. If you give them a map with holes, they can fill in the missing parts based on what real rocks usually look like.
• The Magic: They use a "Diffusion Model." Think of this as starting with a cloud of static noise and slowly cleaning it up until a clear, realistic rock map emerges.

2. The Physicist (The Neural Operator Surrogate)

• Role: This expert is a super-fast simulator. They know the laws of physics: "If the rock looks like this, the gas will flow like that."
• How they work: They are a "surrogate," meaning they are a cheap, fast copy of the real, expensive physics equations. They can instantly tell you what happens if you change the rock map.

How They Work Together (The "Decoupled" Dance)

In the old "Joint" models, the AI tried to learn both the rock shapes and the gas flow at the same time. This often led to the "fizzy static" problem mentioned earlier.

Fun-DDPS separates them:

1. Generation: The Geologist generates a candidate rock map.
2. Testing: The Physicist instantly simulates what happens if that rock map is real.
3. Correction: If the simulation doesn't match the few data points we actually have (the "tiny shards"), the Physicist sends a signal back to the Geologist: "Hey, your rock map is close, but the gas flow doesn't match our measurements. Tweak the rocks slightly."
4. Repeat: The Geologist adjusts the map, and they try again.

Because they are separate, the Geologist never forgets what "real rocks" look like, and the Physicist ensures the laws of physics are never broken.

Why This is a Big Deal (The Results)

1. Handling Extreme Scarcity (The "Blindfold" Test)
The researchers tested this with only 25% of the data available (like having a map with 75% of it torn off).

• Old AI: Failed miserably, making errors of nearly 87%. It was like guessing the shape of a car by looking at a single wheel.
• Fun-DDPS: Succeeded with only 7.7% error. It used its knowledge of what rocks usually look like to fill in the missing 75% perfectly. It was 11 times better than the old methods.

2. No More "Fizzy Static" (Physical Consistency)
When the old "Joint" models tried to guess the underground structure, they produced maps that looked grainy and unnatural (high-frequency artifacts). It was like a photo that was too zoomed in and pixelated.

• Fun-DDPS produced smooth, realistic maps that looked like real geological formations. It avoided the "fizzy static" because the Geologist expert was strictly in charge of the rock shapes, while the Physicist just checked the math.

3. Speed and Accuracy (The "Rejection Sampling" Benchmark)
To prove they were right, they compared their AI to the "Gold Standard" method called Rejection Sampling.

• The Gold Standard: Imagine trying to find a specific needle in a haystack by randomly pulling out 2 million needles and checking each one. It's accurate, but it takes forever (computationally expensive).
• Fun-DDPS: It found the same answer as the Gold Standard but did it 4 times faster. It was like using a metal detector instead of pulling out needles one by one.

The Bottom Line

Fun-DDPS is a smarter way to use AI for underground science. Instead of forcing one AI to be a geologist, a physicist, and a guesser all at once, it splits the job. One AI learns what the world looks like, and another AI checks if the physics works. This allows them to solve incredibly difficult puzzles with very little data, making Carbon Capture and Storage safer and more reliable.

Technical Summary

1. Problem Statement

The paper addresses critical challenges in Carbon Capture and Storage (CCS) regarding the characterization of subsurface flow. Two primary computational tasks are identified:

• Forward Modeling: Forecasting CO2 plume migration and pressure buildup.
• Inverse Modeling: Characterizing subsurface geological heterogeneity (e.g., permeability) from sparse monitoring data.

Key Challenges:

• Ill-posedness & Sparsity: Inverse problems are ill-posed due to extremely sparse observations (often <1% of the domain, e.g., data from only a few wells).
• High Dimensionality & Non-Gaussianity: Subsurface parameters are high-dimensional and exhibit complex, non-Gaussian distributions (e.g., discrete facies, channelized reservoirs) that traditional ensemble methods (like EnKF) fail to capture.
• Computational Cost: Rigorous Bayesian methods (e.g., MCMC) require thousands of high-fidelity simulations, making them infeasible for large-scale 3D models.
• Limitations of Current DL Surrogates: While deep learning surrogates (e.g., FNO) accelerate forward simulations, they are typically deterministic and struggle with inverse problems.
• Failure of Joint-State Diffusion Models: Existing joint-state diffusion models (learning the joint distribution $p(m, s)$ of geology and dynamics) often produce physical inconsistencies and high-frequency artifacts when training data is limited, as they learn statistical correlations rather than explicit physical laws.

2. Methodology: Fun-DDPS

The authors propose Fun-DDPS (Function-space Decoupled Diffusion Posterior Sampling), a generative framework that decouples the learning of geological priors from the approximation of flow physics.

Core Architecture

The framework consists of two independent components:

1. Function-Space Diffusion Prior ($p(m)$):
   • A single-channel diffusion model trained only on geological parameter samples (geomodels, $m$).
   • It learns the prior distribution of geological heterogeneity using a Function-Space Diffusion Model (perturbing fields with Gaussian Random Fields to preserve spatial continuity and discretization invariance).
   • It does not see dynamic state data during training.
2. Differentiable Neural Operator Surrogate ($L_\phi \approx F$):
   • A Local Neural Operator (LNO) trained to approximate the forward physics mapping $F: m \to s$ (from permeability to saturation).
   • The LNO combines global Fourier spectral layers (for pressure responses) with localized kernels (DISCO) to accurately capture sharp saturation fronts without ringing artifacts.

Inference Process (Decoupled Posterior Sampling)

For inverse problems, the method samples from the posterior $p(m|y_{obs})$ by guiding the reverse diffusion process:

• Gradient Guidance: The reverse diffusion step is guided by the gradient of the log-likelihood, $\nabla_m \log p(y_{obs}|m)$.
• Mechanism: Instead of relying on the diffusion model to learn the physics, the pre-trained neural operator surrogate $L_\phi$ is used to compute the likelihood gradient via backpropagation:
  $$\nabla_m \log p(y_{dyn}| \hat{m}_0) \approx -\zeta \nabla_m \| M_{dyn} \odot (L_\phi(\hat{m}_0) - y_{dyn}) \|^2$$
• Benefit: This allows the diffusion prior to robustly recover missing geological features (handling sparsity) while the surrogate provides efficient, physics-consistent guidance to match sparse observations.

3. Key Contributions

1. Decoupled Architecture: The first framework to explicitly separate the learning of geological priors from flow physics in CCS modeling. This avoids the "guidance attenuation" and physical inconsistencies seen in joint-state models.
2. Robustness to Extreme Sparsity: Demonstrates the ability to handle forward modeling with only 25% observation coverage, a regime where deterministic surrogates fail completely.
3. Rigorous Validation: Provides the first rigorous benchmark of diffusion-based inverse solvers against asymptotically exact Rejection Sampling (RS) posteriors, validating statistical accuracy.
4. Physics-Consistent Realizations: Produces geologically realistic samples free from the high-frequency artifacts common in joint-state baselines.

4. Experimental Results

The framework was tested on synthetic CCS datasets (12,000 training pairs) simulating CO2 injection into a deep saline aquifer.

A. Forward Modeling (Partial Geomodel Inputs)

• Setup: Predicting dynamic saturation fields given geomodels with 100%, 50%, and 25% observation coverage.
• Results:
  • Standard Surrogate: Error increased catastrophically from 4.4% (100%) to 86.9% (25%) due to inability to handle missing inputs (zero-filling).
  • Fun-DDPS: Maintained a low relative error of 7.7% even at 25% coverage.
  • Improvement: 11× improvement over standard surrogates.
  • Comparison: Fun-DDPS significantly outperformed the joint-state baseline (Fun-DPS), which showed high errors (34–42%) regardless of sparsity.

B. Inverse Modeling (Sparse Dynamic Observations)

• Setup: Inferring permeability fields from saturation data observed at only two wells (<1% domain coverage).
• Validation: Compared against a "ground truth" posterior generated via Rejection Sampling (RS) using 26,000 accepted samples from 2 million priors.
• Statistical Accuracy:
  • Both Fun-DDPS and Fun-DPS achieved Jensen-Shannon (JS) divergence < 0.06 against the RS reference, indicating high statistical accuracy.
• Physical Consistency:
  • Fun-DPS (Joint): Produced samples with high-frequency, non-physical artifacts (grainy textures) in the posterior mean.
  • Fun-DDPS (Decoupled): Produced smooth, geologically coherent realizations, effectively suppressing noise while honoring data constraints.
• Efficiency: Fun-DDPS achieved these results with 4× improved sample efficiency compared to Rejection Sampling (requiring ~512k evaluations vs. 2M for RS).

5. Significance and Impact

• Solving the "Data Sparsity" Bottleneck: Fun-DDPS proves that generative priors can reconstruct missing geological information effectively, enabling reliable modeling in scenarios where traditional deterministic methods fail.
• Bridging Generative AI and Physics: By decoupling the prior from the physics surrogate, the method ensures that generated samples adhere to physical laws without requiring the generative model to learn complex PDEs from scratch.
• Computational Feasibility: The approach makes rigorous uncertainty quantification feasible for large-scale 3D CCS models by reducing the computational cost by a factor of 4 compared to gold-standard sampling methods, while maintaining physical fidelity.
• Future Directions: The authors note that while the current study uses static snapshots, the decoupled architecture is naturally suited for extension to full spatiotemporal trajectories (time-series data), which is critical for real-world CCS monitoring.

In summary, Fun-DDPS represents a significant advancement in scientific machine learning for subsurface engineering, offering a robust, efficient, and physically consistent solution to the inverse problems inherent in Carbon Capture and Storage.