An integrated experimental and machine learning-based optimization framework for energy-efficient MEA/[TBA][Br] solvent in shipboard CO2 capture

Shipboard CO 2 capture (SCC) systems using MEA/[TBA][Br] solvent lack experimental studies and efficient data-driven optimization frameworks that minimize regeneration energy while maintaining high CO 2 capture rates. This study conducts cyclic experiments and develops a deep neural network (DNN)-based framework to optimize operating conditions for shipboard-relevant carbon capture. In the first stage, 70 experimental samples are collected under varying [TBA][Br] concentration, desorber temperature, and liquid-to-gas (L/G) ratio to train a surrogate DNN model. This model is used to formulate an optimization problem, from which an optimal operating zone is identified. In the second stage, 28 additional experiments are performed within this operating zone, and the full 98-sample dataset is used to train a refined DNN model, featuring three hidden layers with hyperparameters tuned via grid search and evaluated using an 80−20 training-test split. Optimization using the refined DNN identifies the optimal condition at [TBA][Br] concentration of 2.96 wt%, a desorber temperature of 120 C, and an L/G ratio of 1.73 kg/kg, corresponding to more than 90% CO 2 capture and a regeneration energy of 2.94 GJ/t CO 2 . This optimal performance corresponds to an 11.40%–34.87% reduction in regeneration energy compared to baseline MEA in this study and MEA-based SCC systems from the literature. The refined DNN demonstrates strong predictive capability, achieving MAPE values of 1.27% (training) and 1.41% (test) for CO 2 capture rate, and 0.83% (training) and 1.04% (test) for regeneration energy. Sensitivity analysis highlights the influence of all three variables: [TBA][Br] addition reduces sensible and vaporization heat requirements, higher L/G ratios improve CO 2 capture rates, and elevated desorber temperatures simultaneously enhance CO 2 capture rates and lower regeneration energy. Moreover, ATR-FTIR analysis over 15 cycles confirms solvent chemical stability with no evidence of chemical degradation. This work demonstrates an effective machine learning-enabled optimization strategy for SCC with MEA/[TBA][Br] solvent and offers insights into developing energy-efficient SCC systems and a framework extendable to other CO 2 emission sources.