⚡ electrical engineering

Physics-Driven Hierarchical Cascade Surrogate for Compositional Simulation of Geological CO2 and H2S Storage

This study introduces the Hierarchical Cascade Surrogate for CO2 Storage (HCS-CO2S), a physics-driven machine learning model that accurately predicts the full compositional state of reservoir fluids over centennial timescales by decomposing the prediction task into five causally ordered levels to suppress error accumulation in blind forecasts.

Original authors: Stepan Zainulin, Anna Storozheva

Published 2026-07-16✓ Author reviewed ⓘ

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Stepan Zainulin, Anna Storozheva

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Earth's crust as a giant, ancient sponge made of rock, hiding deep underground. Scientists are trying to figure out how to stuff our planet's excess carbon dioxide (CO2) into these deep pockets to stop the climate from warming up. This is called "Geological Carbon Storage." But here's the tricky part: once you pump this gas down there, it doesn't just sit still like a rock. It mixes with salty water, dissolves, floats, and reacts with the rock over hundreds of years. To make sure the gas stays trapped and doesn't leak back up, scientists use super-computers to run complex simulations. These simulations act like a time machine, predicting exactly where the gas will go and how it will change over centuries. However, running these simulations is like trying to solve a million-piece puzzle while the pieces are constantly shifting; it takes so much computing power and time that it's hard to run enough tests to be absolutely sure everything is safe.

Enter the world of "surrogate models." Think of these as clever shortcuts. Instead of solving the entire difficult puzzle from scratch every single time, a surrogate model is like a super-smart student who has watched the master solve the puzzle thousands of times. It learns the patterns and can guess the final picture almost instantly. The big challenge has been making these shortcuts accurate enough to predict not just where the gas is, but exactly what it is doing—whether it's dissolved in water, floating as a gas, or stuck in the rock. If the shortcut gets the details wrong, we might think the gas is safe when it's actually drifting toward the surface.

This paper introduces a new, highly sophisticated shortcut called the "Hierarchical Cascade Surrogate" (HCS-CO2S). Imagine building a house of cards, but instead of just stacking them randomly, you build it in five specific layers, where each layer must be perfectly stable before you can build the next one. The first layer figures out the pressure and temperature (the "weather" inside the rock). The second layer uses that weather to figure out how heavy and thick the fluids are. The third layer decides how much space the gas, oil, and water take up, making sure they add up to exactly 100% of the space available. The fourth layer calculates exactly how much CO2 is dissolved in the water versus floating in the gas. Finally, the fifth layer puts it all together. The researchers trained this model on a simulation of CO2 mixed with a bit of hydrogen sulfide (H2S) being injected into a deep rock formation. They then asked the model to predict what would happen 100 years into the future, a time period it had never seen before. The result? The model successfully predicted the location of the gas cloud and its chemical makeup with impressive accuracy, proving that this "layered" approach can act as a reliable, fast-forward time machine for keeping our planet safe.

The Story of the "Layered" Time Machine

The researchers, Stepan Zainulin and Anna Storozheva, were tackling a problem that feels like trying to predict the path of a drop of ink in a swirling ocean, but the ocean is made of rock and the ink is a toxic gas. They wanted to know: Can we build a machine learning model that doesn't just guess the general shape of the gas cloud, but understands the chemistry inside it?

Most existing models are like taking a blurry photo of the gas cloud. They can tell you roughly where the gas is and how much pressure is building up, but they can't tell you exactly how much of the CO2 has dissolved into the salty water versus how much is still floating as a gas. This is a huge deal because, over hundreds of years, the gas dissolving into the water is actually one of the safest ways to store it. If your model can't see the chemistry, it can't tell you if the storage is truly secure.

The authors argue that previous attempts to speed up these simulations had three main flaws. First, many models only work on perfectly square grids, like a checkerboard, but real underground rock layers are jagged and irregular (like a crumpled piece of paper). Second, some models try to learn the laws of physics directly but often mess up the math over long periods, causing errors to pile up like snowballs rolling down a hill. Third, and most importantly, none of the previous models could predict the detailed chemical breakdown of the fluids.

To fix this, the team built their "Hierarchical Cascade Surrogate." Let's break down how this "cascade" works using a simple analogy. Imagine you are trying to describe a complex dish to a friend. You wouldn't just say "it's tasty." You would follow a logical order:

The Kitchen Conditions (Level 1 & 2): First, you describe the temperature and pressure in the kitchen. Is it hot? Is it high pressure?
The Ingredients' Properties (Level 3): Based on that heat and pressure, you describe how the ingredients behave. Is the oil thick? Is the water thin?
The Mixing (Level 4): Now you figure out how the ingredients fill the bowl. How much space does the gas take? How much does the water take? Crucially, you make sure they add up to exactly one full bowl.
The Flavor Profile (Level 5): Finally, you describe exactly how much of the main spice (CO2) is in the water, how much is in the oil, and how much is in the gas.

This step-by-step approach is the "cascade." By forcing the model to solve the easy, physical parts first (pressure and density) before guessing the complex chemical parts, the model is less likely to make silly mistakes. It's like a chef who checks the oven temperature before deciding how long to bake the cake.

The Big Test: The Blind 100-Year Forecast

The real magic of this paper isn't just that they built the model; it's how they tested it. Usually, scientists train a model on data from years 1 to 100 and then test it on years 101 to 110. But that's cheating a little bit because the model has already seen the pattern.

Instead, the authors did something much harder. They trained the model on data up to a certain point, and then completely removed the final 20 years of data from its training. They then asked the model to predict the state of the reservoir at a specific future time (step 144) based only on an earlier time (step 123). This is a "blind forecast." It's like asking a student to take a final exam on material they haven't seen yet, just to see if they truly understand the concepts.

The results were impressive. The model successfully predicted the location of the CO2 plume (the gas cloud) with near-perfect accuracy.

The Plume Hunt: If you asked the model, "Is there gas in this specific rock cell?" it said "Yes" for every single cell that actually had gas. It didn't miss a single one (100% recall). It did guess "Yes" for 12 extra cells that were actually empty, but that's a "conservative" mistake—it's better to think there's a little extra gas than to miss a dangerous leak.
The Chemistry Check: The model predicted the amount of CO2 dissolved in the water with a high degree of accuracy (an R-squared value of 0.905). This means it could tell the difference between gas that is just floating and gas that is safely locked away in the water.
The Speed: The most exciting part? The full physics simulation that this model is trying to replace takes a long time to run. This new model, running on a standard computer processor (no fancy super-computer needed), predicted the entire future state of the 21,142 rock cells in just 22.4 seconds.

The Flaws and the Future

The authors are honest about where their model isn't perfect. While it was great at predicting the gas cloud's location, it wasn't perfect at predicting the exact pressure in every single spot. In a few specific areas, the pressure prediction was off by up to 11.3 bar. The paper suggests this happens because the model looks at each rock cell individually and doesn't "talk" to its neighbors enough to smooth out the pressure changes, especially after the gas injection stops. It's like a crowd of people where everyone knows their own mood but doesn't know what the person next to them is feeling.

Also, the model was slightly too cautious. It predicted the gas cloud to be a tiny bit smaller than it actually was inside the cloud, but it also thought the cloud extended a little further into the empty rock around it. This is a trade-off the authors made to ensure they didn't miss any gas.

The paper concludes that this "layered" approach is a major step forward. It proves that we can build fast, accurate models that understand the chemistry of carbon storage, not just the shape of the gas cloud. However, the authors are careful to say this was tested on just one type of rock formation. They haven't proven yet that this model works on every different kind of underground rock in the world. That's the next challenge: teaching this smart student to handle different "classrooms" (different geological sites) without needing to relearn everything from scratch.

In the end, this paper doesn't claim to have solved the entire problem of carbon storage. Instead, it offers a powerful new tool—a fast, chemistry-aware time machine—that could help scientists run thousands of safety tests in the time it used to take to run just one. And in the race to save the climate, having more time to check our work might just be the most important thing of all.

Technical Summary: Physics-Driven Hierarchical Cascade Surrogate for Compositional Simulation of Geological CO2 and H2S Storage

Problem Statement
Accurate long-term prediction of CO2 plume migration and inter-phase compositional redistribution is a critical challenge in Carbon Capture and Storage (CCS). While machine learning (ML) surrogates have been developed to accelerate full-physics simulations, existing approaches suffer from four primary limitations:

Grid Incompatibility: Architectures like CNNs and U-Nets require regular Cartesian grids, whereas industrial reservoir modeling relies on stratigraphic corner-point grids with geometric discontinuities.
Physical Consistency Issues: Physics-informed neural networks (PINNs) often fail to guarantee integral mass conservation over long forecast horizons due to singular gradients at sharp saturation fronts.
Lack of Compositional Resolution: Current spatial surrogates typically predict only bulk phase saturations ( $S_g, S_o, S_w$ ) and pressure ( $P$ ), failing to resolve specific molar fractions of CO2 in different phases (gas, oil, water). This prevents direct quantification of solubility trapping, a key long-term storage mechanism.
Temporal Extrapolation: Most surrogates are validated only within their training time windows. Blind forecasting over centennial scales—required for regulatory certification—has not been demonstrated, leading to potential systematic inventory errors as trapping mechanisms shift from structural to solubility/mineral trapping.

Methodology
The authors present the Hierarchical Cascade Surrogate for CO2 Storage (HCS-CO2S), a block-based multi-layer perceptron (MLP) designed for compositional CO2/H2S co-injection on corner-point grids. The methodology is structured around a physically ordered, five-level cascade that mirrors the causal logic of compositional flow solvers:

Feature Construction: The input vector combines static petrophysical properties (porosity, permeability, geometry), dynamic source-step states, and explicit well-influence descriptors. These descriptors (e.g., normalized distance to injectors, Gaussian influence, injection schedule flags) provide implicit spatial embedding for the point-wise MLP, allowing it to learn large-scale plume dynamics without full-field image inputs.
Hierarchical Cascade Architecture:
- Level 1 (Encoder): A shared residual encoder compresses the feature vector into a latent state.
- Level 2 (Thermobaric Head): Predicts reservoir pressure (and temperature if available).
- Level 3 (PVT Head): Predicts phase densities and viscosities conditioned on the thermobaric state.
- Level 4 (Saturation Head): Predicts phase saturations ( $S_g, S_o, S_w$ ) using a gas-gated construction. This enforces the unit-sum constraint ( $S_g + S_o + S_w = 1$ ) exactly by splitting the residual liquid volume via a softmax function.
- Level 5 (Composition Head): Predicts CO2 partitioning (overall mole fraction $z_{CO2}$ , dissolved $R_{sw}$ , liquid $x_{CO2}$ , and gas $y_{CO2}$ fractions). This level applies hard phase masks, ensuring components are only predicted if the corresponding phase is present (e.g., $y_{CO2}$ is zero if $S_g < \theta$ ).
Training and Validation: The model is trained on a high-fidelity tNavigator simulation of a 31×31×22 corner-point grid (21,142 active cells) involving CO2/H2S co-injection. A blind validation protocol is employed where the terminal step pair (steps 123 to 144, representing a 100-year horizon post-injection) is entirely excluded from training to test long-term extrapolation capabilities.

Key Results
Evaluated on the blind 100-year forecast, the HCS-CO2S surrogate demonstrated the following performance:

Physical Consistency: The model strictly adhered to saturation closure ( $S_g + S_o + S_w = 1$ ) with a mean absolute residual of $2.04 \times 10^{-8}$ .
Plume Localization: The surrogate achieved perfect recall (1.000) in detecting gas-bearing cells ( $S_g > 0.05$ ), capturing all 72 true gas cells with an F1 score of 0.923. The predicted plume centroid was displaced by only 1.16 cell-widths from the true centroid.
Compositional Accuracy: The model successfully predicted the overall CO2 mole fraction ( $R^2 = 0.913$ ) and phase-resident fractions ( $R^2 > 0.90$ for dissolved and liquid phases).
Pressure and PVT: Reservoir pressure was reconstructed with an $R^2$ of 0.827 (MAE = 0.798 bar). PVT properties (densities and viscosities) were predicted with high accuracy ( $R^2 > 0.91$ for gas/oil phases).
Limitations Observed:
- Pressure Relaxation: The model exhibited a heavy-tailed residual in pressure (max error 11.3 bar) in cells affected by post-injection relaxation, attributed to the lack of spatial diffusion constraints in the point-wise MLP.
- Saturation Bias: Inside the plume, gas saturation was slightly underpredicted (bias of -0.025), while the plume footprint was mildly over-detected (12 false positives). This conservative behavior is a trade-off for high specificity in the gas gate.
- Metric Interpretation: The low $R^2$ for gas-phase CO2 fraction ( $y_{CO2}$ ) and water viscosity was attributed to data sparsity and near-constant field values, respectively, rather than systematic model bias.

Significance and Claims
The paper claims that HCS-CO2S advances the state of the art in three specific ways:

Compositional Fidelity: It is the first spatial surrogate to resolve full inter-phase compositional states (molar fractions) rather than just bulk saturations, enabling direct assessment of solubility trapping.
Grid and Architecture Compatibility: By using a point-wise MLP with engineered well-influence features, it operates natively on unstructured corner-point grids without requiring upscaling or remeshing, while the hierarchical cascade ensures physical consistency (mass conservation and phase masks) that standard multi-output regressors lack.
Long-Horizon Blind Forecasting: The study demonstrates, for the first time in this domain, a blind forecast over a centennial horizon (100 years post-injection) that successfully reproduces the spatial distribution of compositional targets, addressing a critical regulatory requirement for storage site certification.

The authors conclude that while the surrogate is not yet transferable across different geological realizations (a future direction), it provides a computationally efficient (22.4s on CPU for the full field) and physically consistent tool for uncertainty quantification and screening in compositional CCS modeling.

The Story of the "Layered" Time Machine

The Big Test: The Blind 100-Year Forecast

The Flaws and the Future

Technical Summary: Physics-Driven Hierarchical Cascade Surrogate for Compositional Simulation of Geological CO2 and H2S Storage

More like this