New AI Framework for Carbon Storage Solves Critical Subsurface Modeling Challenge
A novel generative AI framework promises to revolutionize the characterization of underground geology for Carbon Capture and Storage (CCS) projects. The method, called Function-space Decoupled Diffusion Posterior Sampling (Fun-DDPS), tackles the notoriously ill-posed problem of inverse modeling with sparse observational data by combining a function-space diffusion model with a differentiable neural operator surrogate. This decoupled approach enables highly accurate predictions of subsurface flow, a critical factor for ensuring the safe and permanent sequestration of carbon dioxide.
Accurately mapping the underground rock properties, or geomodels, that govern how fluids like CO₂ move is essential for CCS site selection and monitoring. Traditional methods struggle when data from wells and sensors is extremely limited, leading to high uncertainty. The new framework directly addresses this sparsity challenge, offering a path to more reliable and efficient subsurface characterization.
How Fun-DDPS Works: Decoupling Geology from Physics
The core innovation of Fun-DDPS lies in its two-stage, decoupled architecture. First, a single-channel diffusion model is trained to learn a prior probability distribution over possible geological parameter fields. This model becomes adept at generating realistic, spatially coherent geomodels from noise. Second, a Local Neural Operator (LNO) acts as a fast, differentiable surrogate for the complex physics equations governing subsurface flow.
During the inverse modeling process—where the goal is to infer the hidden geology from sparse surface observations—these two components work in concert. The diffusion prior robustly fills in missing geological information, while the LNO surrogate provides efficient, gradient-based guidance to ensure the final model predictions match the observed dynamic data, such as pressure changes. This separation allows each AI component to specialize, leading to superior performance.
Breakthrough Performance on Synthetic CCS Data
Researchers validated Fun-DDPS on synthetic datasets designed to mimic real-world CCS scenarios. The results demonstrate a dramatic leap in capability, particularly in data-sparse environments. For forward modeling tasks with only 25% of observational data available, Fun-DDPS achieved a remarkably low relative error of 7.7%. This stands in stark contrast to the 86.9% error from standard neural surrogate methods, representing an 11x improvement and proving the method's strength where deterministic approaches fail.
Perhaps more significantly, the study provides the first rigorous, quantitative validation of a diffusion-based inverse solver against a gold-standard benchmark. The team compared Fun-DDPS outputs to asymptotically exact posteriors generated by computationally expensive Rejection Sampling (RS). Both Fun-DDPS and a joint-state baseline achieved a Jensen-Shannon divergence of less than 0.06, confirming their statistical accuracy.
Key Advantages: Physical Consistency and Efficiency
Beyond statistical accuracy, Fun-DDPS offers two crucial practical advantages. First, it generates physically consistent realizations of the subsurface. Unlike some joint-state models that can produce unrealistic, high-frequency artifacts, Fun-DDPS's decoupled design ensures the outputs honor the underlying geological and physical principles. Second, it is highly efficient. The framework achieves its high-fidelity results with a 4x improvement in sample efficiency compared to the brute-force Rejection Sampling method, making high-resolution probabilistic modeling computationally feasible.
Why This Matters for Climate Goals
- Enables Safe CCS Deployment: Accurate subsurface models are non-negotiable for verifying that injected CO₂ remains contained and does not leak. This AI framework reduces critical uncertainty.
- Solves the Data Sparsity Problem: It delivers reliable predictions even with very few wells or sensors, lowering the cost and increasing the feasibility of site characterization.
- Provides Probabilistic Forecasts: By generating multiple plausible realizations, it gives engineers a clear picture of risks and uncertainties, supporting better decision-making.
- Bridges AI and Physics: The method successfully integrates data-driven generative AI with physics-based constraints, setting a new standard for scientific machine learning in geoscience.
The development of Fun-DDPS, detailed in the preprint (arXiv:2602.12274v2), marks a significant step toward making large-scale carbon sequestration a more predictable and safer endeavor, directly supporting global climate mitigation efforts.