Hybrid Mechanistic-Machine Learning Surrogate Modelling and Multi-Objective Optimization of Ca(OH)2 - Based Direct Air Capture Systems

Abstract

To mitigate global warming, this research investigates the Direct Air Capture (DAC) of CO2 using calcium hydroxide derived from eggshell waste. A significant barrier to DAC implementation is the high energy cost of implementation and the computational expenses of traditional simulations of the process. This study presents a hybrid mechanistic-machine learning surrogate modelling and optimization framework to identify optimal capture conditions while minimizing energy consumption. A heterogeneous co-simulation framework was built using Aspen Plus for steady state simulations and MATLAB for a pellet-based shrinking core modelling. A Python script integrated these platforms via COM and the MATLAB Engine, using Latin Hypercube Sampling to automate 1000+ simulations across varying temperatures, column height, humidity, pellet sizes, flow rates, etc. The dataset obtained was used to train XGBoost surrogate models, which were then used to carry out Non-Dominated Genetic Sorting optimization. Multiple optimal DAC points with the best trade-offs were identified. One of such points has input variables: CaCO3 flow rate of 97.92kmol/hr, pellet moisture content of 0.065, air flow rate of 71.497kmol/hr, DAC column height of 9.99m, a pellet size of 1.0598mm, air humidity of 0.872, a calcination temperature of 919.69°C and ambient air temperature of 43.71°C. These values gave a DAC efficiency of 96.5% at a specific energy cost of 27.7 dollars per kg of CO2 captured. The surrogate models provided results in seconds, with high predictive power (R2 ≥0.95), demonstrating significant computational superiority over traditional methods and providing a fast, scalable tool for DAC process optimization.