Physicochemical Feature-Driven Machine Learning and Multi-Objective Optimization for CO2 Capture in MEA/PZ Blends

The post-combustion carbon capture process with monoethanolamine/piperazine (MEA/PZ) blends encounters notable modeling and optimization challenges. These arise from strong thermodynamic–kinetic nonlinear coupling, as well as limited availability of high-quality experimental data. To address this, we propose a machine learning and multi-objective optimization strategy driven by physicochemical features. By extracting explicit physical features and embedding physicochemical constraints into data-driven models, this study evaluated the predictive performance of three distinct algorithms based on wet-wall column experimental data. These algorithms included natural gradient boosting (NGBoost), sure independence screening and sparsifying operator (SISSO), and gaussian process regression (GPR). Subsequently, an optimization problem aimed at minimizing PCO2* and maximizing kg′ was formulated. The multi-objective beluga whale optimization (MOBWO) algorithm was then employed for global optimization and benchmarked against the traditional non-dominated sorting genetic algorithm II (NSGA-II). Results indicate that the Gaussian process regression (GPR) model performed best when it was enhanced by physicochemical features and optimized via Bayesian hyperparameter tuning. It achieved R2 values of 0.989 and 0.953 for PCO2* and kg′, with average absolute relative deviations (AARDs) kept below 15.7% and 12.2% respectively. Feature importance analysis validated the underlying physical laws. Specifically, temperature dictates thermodynamic equilibrium, while CO2 loading limits mass transfer kinetics. In the optimization phase, MOBWO outperformed NSGA-II by generating a more uniformly distributed Pareto front. Decision-making analysis further identified three typical operating regimes encompassing kinetics-dominant, thermodynamics-dominant, and comprehensive equilibrium conditions. This framework provides a robust paradigm for small-sample modeling and optimization in complex chemical processes.