Accurately predicting transport properties of porous fibrous materials by machine learning methods

Machine learning algorithms trained on data gathered from stochastically generated gas diffusion layers (GDLs) were used to predict key transport properties that govern effective mass transport behaviour in polymer electrolyte membrane fuel cells. Specifically, we present the largest database in the present literature of stochastically generated fibrous GDL substrates (containing over 2000 unique materials) and the associated structural and transport properties determined via pore network modelling. Seven established machine learning algorithms were trained to predict the effective single-phase permeability (ksp) and diffusivity (Dsp), and the relative permeability (kr) and diffusivity (Dr) of the generated materials using well-defined material properties as input features. Gradient boosting regression (GBR), artificial neural network, and support vector regression were the best performing predictors of the single-phase properties, all of which exhibited statistically insignificant differences in error. GBR provided the best prediction accuracy of relative transport properties. Physical features from a set of generated porous fibrous materials were used to train seven machine learning algorithms to accurately predict the material diffusivity and permeability. The best performing models were found to be gradient boosting regression, artificial neural networks, and support vector regression.image