NeuraCell: a neural-network-based tool for the structural characterization of polymeric porous materials

Jorge Torre-Ordás1,2,3*, Suset Barroso-Solares1,2,3, M. A. Rodríguez-Pérez1,2, Javier Pinto1,2,3

1Cellular Materials Laboratory (CellMat), Condensed Matter Physics, Crystallography, and Mineralogy Department, Faculty of Science, University of Valladolid (Spain)

2BioEcoUVA Research Institute on Bioeconomy, University of Valladolid (Spain)

3Study, Preservation, and Recovery of Archaeological, Historical and Environmental Heritage (AHMAT) Research Group, Condensed Matter Physics, Crystallography, and Mineralogy Department, Faculty of Science, University of Valladolid (Spain)

*jorge.torre@uva.es

The structural characterization is an essential task in the study of porous materials. To achieve reliable results, it requires to evaluate images with hundreds of pores. Current methods require large time amounts and are subjected to human errors and subjectivity. A completely automatic tool would not only speed up the process but also enhance its reliability and reproducibility.

Therefore, the main objective of this article is the study of a deep-learning-based technique for the structural characterization of porous materials, through the use of a convolutional neural network. Several fine-tuned Mask R-CNN models were evaluated using different training configurations in four separate datasets each composed of numerous SEM images of diverse polymeric porous materials: closed-pore extruded polystyrene (XPS), polyurethane (PU), and poly(methyl methacrylate) (PMMA), and open-pore PU.

Concerning the results for the closed-pore materials, the performance evaluations showed that a combined model, trained with all sets of materials at the same time, showcased outstanding performance, overcoming individually trained models. With this combined model, it was demonstrated that NeuraCell can provide highly reliable results in characterizing the pore size, anisotropy ratio, and maximum anisotropy angles, with outcomes equivalent to those of the manual overlay method. The average computation time is estimated to be 7 s per image, making this tool around 515 times faster than the conventional method.

Regarding the evaluation of the open-pore PU trained model, it is also seen that the tool demonstrates remarkable capabilities in accurately characterizing the pore window structures. Despite occasional challenges posed by intersecting pore struts, the method achieves accurate masks and predicts large numbers of pore windows, again resembling the manual method.

All in all, this study demonstrates that deep learning offers a promising future for the accurate and efficient characterization of porous polymeric materials. Despite the complexity of the task, existing models can effectively characterize diverse porous structures, providing researchers with accurate results in a matter of seconds.

References

[1] J. Torre, S. Barroso-Solares, M.A. Rodríguez-Pérez, J. Pinto; Polymer, 2024, 291, 126597.