• Explainable white-box ML predicts CO₂ fugacity coefficient across wide conditions • High accuracy achieved with R² = 0.997 and RMSE = 0.0232 • Closed-form equation extracted from the trained neural network • Sensitivity and leverage analyses confirm robustness and thermodynamic consistency CO 2 fugacity coefficient (ϕ) is an indispensable factor in phase equilibrium computation and process modelling. While this parameter is conventionally calculated using equation of state (EOS) models, the process is computationally intensive and numerically demanding. Contemporary studies have probed machine learning (ML) options; however, documented reports are scarce while extant models are mostly not transparent, non-replicable and not verified for their generalizability. In this investigation, a feed forward neural network that learns via the Levenberg-Marquardt paradigm was built to forecast ϕ at varying pressures and temperatures. To execute this, 640 data records aggregated from experimental trials encompassing an expansive range of thermodynamic conditions was used. The model diagnostics indicate that the model has a high prognostic utility given the following benchmarks: R 2 of 0.997, MSE of 0.00054 and RMSE of 0.0232. To accentuate the interpretability of the model, an exact mathematical formula for ϕ is provided. To further illuminate the model for interpretation purposes, the relative influence of each input was appraised via sensitivity analysis using the relevancy factor. This investigation exposes temperature as the leading factor with a percentage contribution of 64% while pressure contributes 36%. Furthermore, trend analysis established the model to be in concordance with the thermodynamics of CO 2 fugacity while the leverage plot diagnostics confirm that 99.5% of the entries in the database lie within the prognostication valid domain. Given that the model is interpretable, is presented in a closed equation form and it exhibits reasonable forecasting precision, then it can be used as a reliable substitute for the rigorous and cumbersome EOS models for ϕ estimation.
Agwu et al. (Fri,) studied this question.