A Regression-Based Predictive Model Hierarchy for Nonwoven Tensile Strength Inference

Dario Antweiler, Jan Pablo Burgard, Marc Harmening, Nicole Marheineke, Andre Schmeißer, Raimund Wegener, Pascal WelkeCognitive 2025 · DOI · OpenAlex

Abstract

Nonwoven materials, characterized by a random fiber structure, are essential for various applications including insulation and filtering. An industrial long-term goal is to establish a framework for the simulation-based design of nonwovens. Due to the random structures, simulations of material properties on fiber network level are computational expensive. We propose a predictive model hierarchy for inferring an important material property—the nonwoven tensile strength behavior. The model hierarchy is built using regression-based approaches, including linear and polynomial models, which provide interpretable results. This allows for significant speedup (six orders of magnitude) over the conventional simulations, while achieving good prediction results ($R^2=0.95$). The proposed models open the application to nonwoven material design, as they provide accurate and cost-effective surrogates for predicting material properties. In this way, our work serves as a proof of concept.

Introduction

Nonwoven materials—ubiquitous in products ranging from N95 masks and air filters to automotive insulation—are manufactured by randomly depositing fibers and bonding them together. For manufacturers, the ability to predict how adjustments in production parameters (such as conveyor speed or bonding strength) affect the final material's physical properties is critical for efficient design.

Overview of the nonwoven material lifecycle: (a) simulation of fiber deposition, (b) industrial production line, (c) resulting material sample, and (d) physical tensile testing.

Traditionally, predicting these properties requires heavy physics-based simulations. Because the fiber structure is inherently random, researchers must generate multiple virtual fiber networks and run complex differential equations on each (Monte-Carlo simulations) to determine average strength. This process is computationally intensive, often requiring 24 to 48 hours to simulate a single sample, creating a significant bottleneck for iterative material design.

Methodology: The Predictive Model Hierarchy

To bypass these computational limits, this paper introduces a "predictive model hierarchy" using Informed Machine Learning. The authors developed three regression-based surrogate models to replace the slow physics simulators:

The proposed predictive model hierarchy, illustrating the relationships between production parameters (u), fiber graphs (v), and material behavior (y).
  1. PP-model (Production Parameter-based): A basic linear regression mapping production parameters directly to stress-strain curves. While fast, it lacks accuracy due to the complex, non-linear physical interactions involved.
  2. FGF-model (Fiber Graph Feature-based): This model extracts topological features from a virtual fiber graph (like shortest paths and node degrees) to predict strength. It is highly accurate but still requires the time-consuming step of generating random graphs.
  3. PP-MGF-model (The Novel Contribution): This sequential model offers a balance of speed and accuracy. It uses multivariate polynomial regression to predict mean graph features directly from production parameters, then feeds those features into an errors-in-variables linear model to predict the final stress-strain curve. This approach completely skips expensive graph generation and physics simulation during inference.

Data Strategy

The models were trained using a combination of labeled and unlabeled data generated by the physics-based TSS-model. Labeled data (Sets 1 & 2) provided ground truth for 187 fully simulated curves, while a larger unlabeled dataset (Set 3) of 2,000 fiber graphs allowed for robust training of the intermediate polynomial regression models.

Key Results and Insights

The research yielded several critical insights into the relationship between fiber topology and material strength:

  • Curve Parameterization: The complex stress-strain behavior can be summarized by two parameters: a threshold strain and a quadratic growth coefficient.
  • Topology as a Predictor: Topological features of the fiber graphs strongly encode the material's physical strength. By splitting the prediction into two steps—parameters to topology, and topology to strength—the authors achieved both interpretability and speed.
  • Massive Speedup: The PP-MGF-model reduces inference time from days to just 10 milliseconds per sample, representing a speedup of over six orders of magnitude.
  • High Accuracy: Using a degree-5 polynomial, the PP-MGF-model achieved a median $R^2$ of 0.9584, nearly matching the performance of the much slower physics-dependent models.

Conclusion

This predictive model hierarchy serves as a successful proof of concept for the simulation-based design of nonwoven materials. By replacing expensive differential equation solvers with efficient regression surrogates, the framework enables the rapid optimization of production parameters that was previously computationally impossible. Future work will focus on deploying these models in real-world industrial design environments.

Summary updated

Abstract

Abstract Nonwoven materials, characterized by a random fiber structure, are essential for various applications including insulation and filtering. An industrial long-term goal is to establish a framework for the simulation-based design of nonwovens. Due to the random structures, simulations of material properties on fiber network level are computational expensive. We propose a predictive model hierarchy for inferring an important material property—the nonwoven tensile strength behavior. The model hierarchy is built using regression-based approaches, including linear and polynomial models, which provide interpretable results. This allows for significant speedup (six orders of magnitude) over the conventional simulations, while achieving good prediction results ( $$R^2=0.95$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msup> <mml:mrow> <mml:mi>R</mml:mi> </mml:mrow> <mml:mrow> <mml:mn>2</mml:mn> </mml:mrow> </mml:msup> <mml:mo>=</mml:mo> <mml:mn>0.95</mml:mn> </mml:math> ). The proposed models open the application to nonwoven material design, as they provide accurate and cost-effective surrogates for predicting material properties. In this way, our work serves as a proof of concept.
BibTeX
@inproceedings{antweiler2025regression,
  author = {Dario Antweiler and Jan Pablo Burgard and Marc Harmening and Nicole Marheineke and Andre Schmeißer and Raimund Wegener and Pascal Welke},
  title = {{A Regression-Based Predictive Model Hierarchy for Nonwoven Tensile Strength Inference}},
  booktitle = {Cognitive},
  year = {2025},
  doi = {10.1007/978-3-031-83097-6\_4},
}
Figure