Multivariate property analysis of CO2 adsorption in South African coals using machine learning based modelling

• Machine learning accurately predict CO 2 adsorption in South African coals. • Volatile matter, fixed carbon, and vitrinite strongly affect adsorption behaviour. • GPR and ANN models show high accuracy in isotherm prediction. • Coal characteristics linked via empirical multivariate models. Accurate prediction of CO 2 adsorption in coal seams is critical for carbon capture and storage (CCS) applications, yet traditional isotherm models require sample-specific fitting and extensive laboratory testing. This article reports the empirical and multi-variable models to predict CO 2 adsorption isotherms in South African coals using proximate, ultimate and petrographic characteristics combined with pressure as input variables. Ten coal samples of varying rank were analysed, and their adsorption behaviour was modelled using neural networks, Gaussian Process Regression (GPR), and Binary Decision Trees (BDT). The models achieved high predictive accuracy (R 2 > 0.98) with low error metrics, illustrating the possibility of achieving a general isotherm prediction without individual fitting. Two analysis techniques (multivariate analysis) revealed strong interrelationships among coal characteristics, enabling empirical models for estimating key properties. All models were trained on data obtained from conventional isotherm fitting, along with predictive coal characteristics values. The single-layer network and GBR models provided satisfactory accuracy without overfitting and lower complexity.