Comparative study of machine learning techniques for post-combustion carbon capture systems.

Area of Science:

• Chemical Engineering
• Computational Science
• Machine Learning

Background:

• Computational fluid dynamics (CFD) simulations are crucial for analyzing packed absorption columns in carbon capture systems (CCSs).
• CFD models capture complex interactions and mass transfer but are computationally intensive, hindering industrial-scale design optimization.
• Accelerating the evaluation of various designs and operating conditions is vital for improving CCS efficiency.

Purpose of the Study:

• To explore the application of machine learning (ML) methods, specifically convolutional neural networks (CNNs) and graph neural networks (GNNs), for aiding and accelerating CCS design.
• To train ML models using CFD datasets to predict key CO2 capture efficiency determinants.
• To evaluate the impact of different input features on model accuracy and generalizability.

Main Methods:

• Utilized statistical ML methods, CNNs, and GNNs on existing CFD datasets of countercurrent flows in packed absorption columns.
• Trained models to predict CO2 capture efficiency using column geometric parameters and inlet velocity conditions.
• Assessed the influence of various input representations on model performance.

Main Results:

• Developed ML models capable of estimating CO2 capture efficiency without requiring additional CFD simulations.
• Demonstrated the potential of CNNs and GNNs to accurately predict performance based on geometric and operational parameters.
• Identified the impact of different input types on the accuracy and generalizability of the ML models.

Conclusions:

• ML approaches, including CNNs and GNNs, offer a computationally efficient alternative to traditional CFD for CCS design and optimization.
• These methods can significantly accelerate the scale-up process for solvent-based post-combustion carbon capture technologies.
• Further research into ML applications can enhance CO2 capture property prediction and streamline the development of more efficient CCS.