Abstract
The CO2 capture efficiency in solvent-based carbon capture systems (CCSs) critically depends on the gas-solvent interfacial area (IA), making maximization of IA a foundational challenge in CCS design. While the IA associated with a particular CCS design can be estimated via a computational fluid dynamics (CFD) simulation, using CFD to derive the IAs associated with numerous CCS designs is prohibitively costly. Fortunately, previous works such as Deep Fluids (DF) show that large simulation speedups are achievable by replacing CFD simulators with neural network (NN) surrogates that faithfully mimic the CFD simulation process. This raises the possibility of a fast, accurate replacement for a CFD simulator and therefore efficient approximation of the IAs required by CCS design optimization. Thus, here, we build on the DF approach to develop surrogates that can successfully be applied to our complex carbon-capture CFD simulations. Our optimized DF-style surrogates produce large speedups (4000x) while obtaining IA relative errors as low as 4% on unseen CCS configurations that lie within the range of training configurations. This hints at the promise of NN surrogates for our CCS design optimization problem. Nonetheless, DF has inherent limitations with respect to CCS design (e.g., limited transferability of trained models to new CCS packings). We conclude with ideas to address these challenges.
Introduction
To combat climate change, efficient Carbon Capture Systems (CCS) are essential. A primary design objective is maximizing the Interfacial Area (IA)— the surface where liquid solvent interacts with flue gas to remove CO2. Traditionally, engineers rely on Computational Fluid Dynamics (CFD) to calculate IA for various equipment configurations. However, CFD is computationally intensive; a single simulation can require hours on high-performance computing clusters, rendering iterative design optimization across thousands of configurations impractical. This research develops a fast neural network surrogate that mimics CFD solvers at a fraction of the computational cost.
Methodology
This study builds upon the Deep Fluids (DF) framework, which utilizes a two-part architecture to simulate physics within a compressed latent space:
- Latent Vector Model (LVM): A Convolutional Neural Network (CNN) autoencoder compresses high-dimensional fluid snapshots into a low-dimensional latent vector.
- Latent Integration Network (LIN): A sequence model (e.g., Transformer or MLP) predicts the temporal evolution of these latent vectors.
Key technical innovations include Parameter Injection, where solvent inlet velocity ($v$) and time ($t$) are directly integrated into the latent code, and End-to-End (E2E) Training, which fine-tunes the compressor and predictor simultaneously to optimize for future state accuracy. The researchers also conducted an architecture search comparing Vision Transformers (ViTs), LSTMs, and Autoregressive CNNs (ARC).
Data and Preprocessing
The dataset comprised 50 CFD simulations of a 2D vertical slice derived from a 3D Representative Column Model (RCM). The raw CFD data, totaling approximately 182GB, was downsampled and interpolated onto a uniform $64 \times 64$ grid to ensure compatibility with standard neural network architectures. The data was split into 40 simulations for training and 10 for testing on unseen inlet velocities.
Results and Insights
The surrogate model achieved a 4000x speedup compared to traditional CFD simulations while maintaining high accuracy, with a 4% relative error in predicting the steady-state gas-solvent interfacial area on unseen configurations.
Several key insights emerged during development:
- Inductive Bias: Standard CNNs outperformed Vision Transformers as autoencoders, likely due to the CNN's inherent spatial locality biases being more effective on the relatively small dataset.
- Loss Function Optimization: Optimizing for Relative Error proved superior to standard Mean Squared Error (RMSE). This approach prevents the model from neglecting small-scale but physically significant features in the fluid dynamics.
- Steady-State Performance: The E2E training significantly improved the model's ability to reach and accurately represent the steady-state phase, which is critical for design optimization.
Discussion and Future Work
While the 2D results are promising, scaling to 3D domains with millions of data points remains a significant hurdle. Furthermore, the current models lack transferability; changing the physical geometry of the "packing" requires retraining the model. Future research will explore Graph Neural Networks (GNNs) to handle varying geometries without retraining, potentially providing a more robust tool for comprehensive CCS design optimization.
Summary updated