AI in Amine Degradation and Carbon Capture Optimization

This article explores how chemical process engineers leverage artificial intelligence to model amine solvent degradation and optimize post-combustion carbon capture absorption columns. As industries strive to reduce greenhouse gas emissions, post-combustion carbon capture using amine solvents remains the leading technology, yet it suffers from high operational costs due to solvent degradation and energy-intensive regeneration. By integrating machine learning algorithms with real-time plant data and thermodynamic principles, engineers can accurately predict chemical degradation pathways, dynamically adjust absorption column parameters, and maximize carbon capture efficiency while minimizing energy consumption.

Understanding the Challenges of Amine-Based Carbon Capture

Post-combustion carbon capture relies heavily on chemical absorption using aqueous amine solvents, such as monoethanolamine (MEA). Flue gas enters the bottom of an absorption column and contacts the descending amine solution, which selectively absorbs carbon dioxide (\(CO_2\)). While effective, this process faces two major hurdles:

  1. Solvent Degradation: Amines degrade over time due to reactions with oxygen (oxidative degradation) and high temperatures in the stripper column (thermal degradation), as well as reactions with flue gas impurities like \(SO_x\) and \(NO_x\). This reduces capture capacity, causes equipment corrosion, and generates toxic byproducts.
  2. Column Mass Transfer Inefficiencies: Fluctuations in power plant load alter flue gas flow rates and \(CO_2\) concentrations. Traditional linear control systems struggle to adjust the complex, non-linear hydrodynamics inside the absorption column quickly enough to maintain optimal efficiency.

Modeling Amine Solvent Degradation with AI

Traditional first-principles models struggle to predict amine degradation because the underlying reaction networks involve dozens of poorly understood intermediate compounds and complex kinetics. Chemical process engineers use AI to bypass these limitations by finding hidden patterns in operational data.

  • Data Collection and Feature Selection: Engineers feed historical plant data into AI models. Key inputs include stripper reboiler temperatures, dissolved oxygen levels, flue gas contaminant concentrations (\(SO_2\), \(NO_2\)), solvent circulation rates, and dissolved metal ion concentrations (which catalyze degradation).
  • Machine Learning Algorithms: Engineers utilize Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) networks because solvent degradation is a time-dependent process. Random Forest and Support Vector Regression (SVR) models are also used to correlate specific operational spikes with sudden increases in heat-stable salts and degradation byproducts like ammonia and формальдегид.
  • Physics-Informed Neural Networks (PINNs): To prevent AI from making chemically impossible predictions, engineers use hybrid models. PINNs embed fundamental laws—such as mass conservation and known reaction kinetics—directly into the neural network’s loss function. This ensures the AI predicts realistic degradation rates even when exposed to new operating conditions.
  • Predictive Maintenance and Reclaiming: Once trained, the AI acts as a "soft sensor." It predicts the exact concentration of degraded solvent in real time without waiting for manual lab titration. Engineers use these predictions to schedule optimal solvent reclaiming or makeup additions, preventing foaming, fouling, and corrosion before they impact the plant.

Optimizing the Absorption Column

The absorption column is a dynamic environment governed by gas-liquid mass transfer, reaction kinetics, and thermodynamics. AI optimizes this equipment by transitioning the plant from reactive control to proactive, predictive control.

  • Surrogate Modeling for Fast Simulation: Rigorous rate-based simulations of absorption columns (using software like Aspen Plus) are computationally heavy and too slow for real-time control. Engineers train deep neural networks on data generated by these simulators. The resulting AI "surrogate model" can predict column temperature profiles and \(CO_2\) capture rates in milliseconds rather than hours.
  • Model Predictive Control (MPC) and Reinforcement Learning: Engineers integrate these fast AI models into the plant’s control architecture. Reinforcement Learning (RL) agents continuously interact with the digital twin of the absorption column, learning the best control policies through trial and error.
  • Dynamic Parameter Adjustment: As incoming flue gas properties fluctuate, the AI-driven control system dynamically adjusts critical variables:
    • Liquid-to-Gas (L/G) Ratio: The AI modulates the lean amine solvent flow rate to match the exact \(CO_2\) load in the flue gas, preventing over-pumping and saving electricity.
    • Intercooling Temperatures: Absorption is an exothermic process. If the column gets too hot, the equilibrium shifts, reducing \(CO_2\) absorption. AI optimizes intercooler duty to maintain the exact temperature profile required for maximum mass transfer rates.
    • Rich Solvent Loading: By maximizing the amount of \(CO_2\) absorbed per unit of solvent exiting the bottom of the column, the AI directly reduces the steam required to regenerate the solvent in the downstream stripper column.

The Operational Impact

By combining degradation modeling with column optimization, chemical process engineers create a unified control strategy. If the optimization AI wants to push the system harder by increasing temperatures, the degradation AI acts as a constraint, warning the system of the financial penalty associated with accelerated solvent loss. This synergy reduces the energy penalty of carbon capture, extends the lifespan of expensive amine solvents, and makes industrial decarbonization economically viable.