Graph-based tensile strength approximation of random nonwoven materials by interpretable regression

Machine learning framework to predict nonwoven material properties from fiber graph representations

“…⃝ a surrogate model for the simulation of the fiber lay-down, 2 ⃝ the generation of a fiber graph, 3 ⃝ a conventional solver for an ordinary differential equation describing the nonwovens' mechanical behavior under vertical load (ODE-solver), 4 ⃝ and the final predicted stress-strain curve of the material. The presented machine learning approach can reliably approximate the resulting curves via regression based on selected graph features while achieving a 1000× speedup.…”

“…The employed tensile strength model-simulation framework (ii), originating from [3], recreates the elastic phase of the nonwovens' tensile strength behavior under vertical load. The suitability of the simulation results is discussed in [4]. Particularly, the model describes the mechanical behavior of the adhered fiber structure by capturing the interaction of the individual fiber connections, each equipped with a nonlinear material law, at network level.…”

“…We approach the problem outlined above by introducing (iii) an interpretable machine learning regression model [4]. Subject of the predictions are the stress-strain curves associated to individual fiber graphs.…”

“…This comprises a selected set of standard graph features, such as the number of nodes and edges or the lengths of (weighted) shortest paths connecting the top to the bottom of the samples, as well as a set of stretch features. The stretch features are determined using a novel stretching algorithm [4] that is based on a reduced model of the nonwovens' tensile behavior. This already encodes a lot of information for predictions, especially with regard to predicting the sample elongation at which fibers straighten out and increased stress occurs.…”

“…If run_computation.sh all is executed, a completely new dataset is generated and used for training and validation. Initial production parameters are sampled from ranges defined in [4], which may be adjusted in the Matlab routine run-DataBaseGeneration.m. It takes the arguments ''NFullyLabeled'', ''NSingleLabeled'' as well as ''NSamples''.…”