Estimation of the three-dimensional permeability tensor of engineering textiles using artificial neural networks

The permeability quantifies the conductance of a textile as a porous medium for liquid flow and is an important material property for the liquid composite moulding (LCM) process. There are many different systems to measure and determine the permeability. The advantage of systems based on 3-dimensional flow progression is that the permeability tensor can be determined with one single measurement. However, derivation of the permeability values from the experimental data can be complex, especially when the flow front reaches the cavity walls and no analytical solution is available. In this study, a new artificial-intelligence-based evaluation method is presented for this case. An artificial neural network (ANN) is trained and tested with data from finite volume method (FVM) simulations. Different architectures and training data are used to evaluate the influences on the quality of the permeability estimation. With a minimum of 345 training data sets, a mean absolute percentage error (MAPE) of approx. 2.57 % was achieved during the training. The training can be further improved, when the simulation data is used multiple times. The trained ANNs are then applied to experimental data. Here, the estimation error for individual cases can be detected by a cross-check of two different ANNs, to improve reliability. A small deviation between the ANNs refers to a valid estimation. The developed method shows a promising approach to estimate the 3-dimensional permeability in a single experiment. It could also be used to extract permeability from flow in geometrically complex, real parts.