💻 computer science

CO $_2$ sequestration hybrid solver using isogeometric alternating-directions and collocation-based robust variational physics informed neural networks (IGA-ADS-CRVPINN)

This paper introduces a hybrid solver combining IsoGeometric Analysis Alternating Directions (IGA-ADS) and Collocation-based Robust Variational Physics Informed Neural Networks (CRVPINN) to efficiently simulate CO₂ sequestration in porous structures, demonstrating a computational speedup of over three times compared to a traditional direct solver baseline.

Original authors: Askold Vilkha, Tomasz Służalec, Marcin Łoś, Maciej Paszyński

Published 2026-04-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Askold Vilkha, Tomasz Służalec, Marcin Łoś, Maciej Paszyński

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

The Big Picture: Trapping Carbon Underground

Imagine the Earth's crust as a giant, giant sponge made of rock. This "sponge" is full of tiny holes (porosity) and has pathways that let water flow through it (permeability).

Scientists want to solve a major problem: too much carbon dioxide (CO2) is in our atmosphere, warming the planet. The solution? CO2 Sequestration. This means capturing the CO2 from factories and pumping it deep underground into these rock sponges to lock it away forever.

But before we pump it in, we need to know: Where will the gas go? How fast will it spread? And how much pressure will it build up? If we guess wrong, the gas might leak back out or crack the rock.

The Problem: Two Different Puzzles

To predict how the gas moves, scientists use math equations. In this paper, the authors are trying to solve two specific puzzles at the same time:

The Saturation Puzzle (Where is the gas?): This changes quickly over time as gas is pumped in. It's like watching a drop of food coloring spread through a glass of water.
The Pressure Puzzle (How hard is it pushing?): This is a "stationary" problem. It doesn't change as fast, but it's mathematically very heavy and difficult to calculate. It's like trying to figure out the exact pressure in a balloon without popping it.

The Old Way: The "Brute Force" Method

Traditionally, to solve the Pressure Puzzle, scientists use a very powerful, old-school calculator called MUMPS.

The Analogy: Imagine you are trying to solve a massive jigsaw puzzle. The MUMPS solver is like a super-intelligent robot that looks at every single piece, compares it to every other piece, and forces them together perfectly.
The Downside: It is incredibly accurate, but it is also slow. It takes a long time to crunch the numbers, especially for big, complex maps of underground rock.

The New Way: The "Hybrid" Super-Solver

The authors of this paper created a new team-up, which they call IGA-ADS-CRVPINN. Think of it as a relay race with two specialized runners who are much faster together than the old robot.

Runner 1: The Saturation Specialist (IGA-ADS)

Who they are: This part of the system is great at tracking the gas spreading (saturation).
How they work: They use a technique called "Isogeometric Analysis."
The Analogy: Imagine drawing a map with smooth, flowing curves (like a painter's brushstroke) instead of jagged, blocky pixels. This allows them to track the gas movement very smoothly and quickly using standard math.

Runner 2: The Pressure Pro (CRVPINN)

Who they are: This is the star of the show. Instead of the slow "brute force" robot, they use a Neural Network (a type of Artificial Intelligence).
How they work: This AI is "Physics-Informed." It doesn't just guess; it knows the laws of physics (like how fluids move) and uses them to learn the answer.
The Analogy: Imagine you want to know the pressure in a balloon.
- The Old Way (MUMPS) is like measuring every single molecule of air inside the balloon one by one.
- The New Way (CRVPINN) is like a seasoned expert who looks at the balloon, feels the tension, and instantly knows the pressure based on experience and rules.
- The "Pre-training" Trick: Before the race starts, the AI is trained for a while (20,000 rounds) to learn the basic rules of the specific underground rock. Once it's trained, it only needs a tiny "tune-up" (100 rounds) for every new step of the simulation.

Why is this a Big Deal?

The authors tested their new hybrid team against the old "brute force" team.

The Result: The new hybrid team was more than 3 times faster.
The Cost: The old method took about 3 hours (10,800 seconds) to run a simulation. The new method took about 57 minutes (3,400 seconds).
The Hardware: The new method can even run on a standard laptop or a free Google Colab account, whereas the old method usually needs a massive, expensive supercomputer.

The "Secret Sauce": Robust Loss

You might ask, "How does the AI know it's right?"
The paper mentions a "Robust Loss" function.

The Analogy: Imagine the AI is a student taking a test. A normal test might give partial credit for being "close." This "Robust" test is smarter. It uses a special scoring system (involving something called a "Gram Matrix") that ensures the AI doesn't just get lucky; it has to be mathematically consistent with the laws of physics to get a good score. This prevents the AI from making wild, wrong guesses.

What's Next?

The authors are excited because this speed-up opens up new possibilities:

Inverse Problems: Instead of asking "Where will the gas go?", they can ask "Where should we put the injection well to get the best result?" (Solving the puzzle backwards).
Hydrogen Storage: They plan to use this same fast method to figure out how to store Hydrogen fuel underground.
Accessibility: Because it's so fast and light, more scientists and engineers can use it without needing a billion-dollar supercomputer.

Summary

This paper presents a speed-boost upgrade for simulating how to store CO2 underground. By swapping a slow, heavy-duty math calculator for a smart, fast AI assistant (Neural Network) that works in a team with a smooth-mapping tool, they made the simulation 3 times faster without losing much accuracy. It's like upgrading from a horse-drawn carriage to a sports car for saving the planet.

1. Problem Statement

The paper addresses the simulation of CO2 sequestration in underground porous structures (e.g., aquifers, abandoned wells). The goal is to model the physical behavior of injected CO2 gas interacting with brine solution, excluding chemical reactions.

The mathematical model is based on Darcy's Law, resulting in a coupled system of two partial differential equations (PDEs):

Saturation Equation: A time-dependent equation describing the evolution of gas saturation ( $S_g$ ).
Pressure Equation: A stationary (elliptic) equation describing the pressure field ( $p$ ), which depends on the current saturation.

Key Challenge: While the time-dependent saturation equation is well-suited for standard numerical methods, the stationary pressure equation is computationally expensive to solve at every time step using traditional direct solvers, creating a bottleneck in the simulation workflow.

2. Methodology: The Hybrid Solver (IGA-ADS-CRVPINN)

The authors propose a hybrid solver that combines two distinct numerical approaches to optimize performance:

A. IsoGeometric Analysis Alternating Directions Solver (IGA-ADS)

Role: Computes the saturation scalar field ( $S_g$ ).
Technique: Uses the Finite Element Method (FEM) with B-spline basis functions (Isogeometric Analysis).
Mechanism: It employs an explicit time-stepping method (Forward Euler) and utilizes Alternating Directions to solve the time-dependent saturation equation. The use of B-splines allows for high-order continuity and efficient $L^2$ projections.
Advantage: The alternating directions approach allows the system to be solved with linear computational cost ( $O(N)$ ) by inverting banded matrices using the Thomas algorithm.

B. Collocation-based Robust Variational Physics Informed Neural Networks (CRVPINN)

Role: Computes the pressure scalar field ( $p$ ).
Technique: A specialized variant of Physics Informed Neural Networks (PINNs).
Mechanism:
- Variational Formulation: Instead of minimizing strong residuals (standard PINN), CRVPINN minimizes weak residuals derived from the integral form of the PDE.
- Collocation & Kronecker Delta: It uses Kronecker delta test functions at collocation points to avoid expensive numerical integration of weak forms.
- Robust Loss Function: To ensure stability (inf-sup stability), the loss function incorporates the inverse of the Gram matrix (constructed from inner products of test functions). This is computed efficiently via LU factorization (performed once) rather than direct inversion at every step.
- Training Strategy: The network is pretrained on the initial pressure configuration. For subsequent time steps, it undergoes a very short training phase (only 100 iterations of the Adam optimizer) using the updated saturation field as input.

C. The Coupling Workflow

Initialization: Pretrain CRVPINN for the initial pressure state.
Time Step Loop:
- Use IGA-ADS to update the saturation field ( $S_g^{n+1}$ ) based on the current pressure ( $p^n$ ).
- Feed the new saturation field into the CRVPINN.
- Run a lightweight CRVPINN training (100 epochs) to update the pressure field ( $p^{n+1}$ ).
- Project the neural network output back to the B-spline basis for the next IGA-ADS step.

3. Key Contributions

Novel Hybrid Architecture: Successfully couples a traditional high-order FEM solver (IGA-ADS) with a modern deep learning solver (CRVPINN) for a multiphase flow problem.
Efficiency via Pretraining: Demonstrates that a neural network can be effectively "warm-started" and updated with minimal iterations (100 steps) per time step, bypassing the need for full retraining.
Robust Loss Function: Implementation of the Gram matrix inverse in the loss function ensures the PINN solver remains stable and accurate for elliptic problems, addressing a common weakness in standard PINNs.
Scalability: The method is designed to run on heterogeneous hardware, utilizing CPUs for the FEM part and GPUs for the neural network part.

4. Results

The study was conducted on the ARES cluster (ACK CYFRONET) using three different reservoir configurations (K1, K2, K3) with varying porosity and permeability maps.

Accuracy:
- Uniform Permeability: The hybrid solver achieved a relative error of <5% compared to the baseline IGA-ADS + MUMPS solver.
- Non-Uniform Permeability: The error increased slightly to 15–20% in complex regions, which the authors attribute to the complexity of the permeability map, though the results remained physically reasonable.
Performance (Speed):
- Baseline (IGA-ADS + MUMPS): Took approximately 10,800 seconds for 1,000 time steps (dominated by the MUMPS direct solver for pressure).
- Hybrid (IGA-ADS + CRVPINN): Took approximately 3,400 seconds for the same task.
- Speedup: The hybrid solver is over 3 times faster than the baseline.
- Breakdown: The CRVPINN training per step took only ~3–5 seconds, whereas the MUMPS solver took significantly longer per step.
Hardware Utilization:
- IGA-ADS ran on a multi-core CPU (Intel Xeon).
- CRVPINN ran on a single NVIDIA A100 GPU.
- The method also demonstrated feasibility on low-resource hardware (laptops) and free cloud platforms (Google Colab).

5. Significance and Future Work

Significance: This work proves that hybridizing classical numerical methods with neural networks can significantly accelerate complex multiphysics simulations without sacrificing too much accuracy. It offers a viable path for real-time or near-real-time reservoir simulation, which is crucial for optimizing CO2 injection strategies.
Future Directions:
- Inverse Problems: Using the solver to identify optimal injection well locations and reservoir parameters.
- Hydrogen Storage: Adapting the framework for H2 storage simulations.
- Optimization: Reducing communication overhead between the CPU and GPU components and further parallelizing the CRVPINN across multiple GPUs.
- Adaptivity: Implementing $hp$-adaptive methods to improve accuracy in complex regions.

In conclusion, the IGA-ADS-CRVPINN solver represents a significant step forward in computational geoscience, offering a fast, robust, and scalable alternative to traditional direct solvers for CO2 sequestration modeling.