Development of machine learning models for the prediction of binary diffusion coefficients of gases

Diffusion coefficient (D₁₂) is an important transport property in the petrochemical and pharmaceutical industries for the design and optimization of processes. The process of measuring the D₁₂ for gas mixtures via an experimental approach is time-consuming and technically challenging. Consequently, many empirical models have been developed to circumvent these challenges. Though these models exhibit good agreement with the experiment, nonetheless, improvement is still needed in terms of increasing the model performance, and the flexibility of their applications because some of these models require extensive calculations to obtain their input parameters. Motivated by this, the work presents a simple and accurate approach for the estimation of diffusion coefficient using machine learning (ML) algorithms. This study employs support vector regression (SVR), Gaussian process regression (GPR), and artificial neural networks (ANN) to estimate the D₁₂ of molecular gas systems over a wide range of temperature (293- 313K), and pressure (0.05–12.0 MPa). The proposed ML models were built using simple descriptors such as temperature, pressure, molar masses of constituent mixtures, and mole fractions of the gases. On the testing dataset, the following overall correlation coefficients were obtained 90.7 %, 99.1 %, and 99.97 % percent for the SVR, GPR, and ANN, respectively. The ANN model did better than the other ML algorithms presented in terms of its generalization capability. The authors believe that the ANN model presented is a viable alternative for the evaluation of the binary diffusion coefficient of gases due to the simplicity of the descriptors.