Benchmarking Machine Learning Models on Treloar's Classic Hyperelasticity Data
A recent preprint on arXiv (2608.14063) evaluates various machine-learning-based constitutive models for hyperelastic materials, utilizing the well-known Treloar experimental dataset. The research contrasts several approaches, including Constitutive Artificial Neural Networks (Generalized-Invariant), Physics-Augmented Neural Networks, Material Fingerprinting (Adaptive), and Efficient Unsupervised Constitutive Law Identification & Discovery. The authors assess aspects such as fitting performance, computational expenses, sensitivity to hyperparameters, and implementation simplicity. They also explore the trade-offs between model complexity and predictive accuracy, measured by the number of material parameters and computational duration. This study highlights the increasing necessity for systematic evaluations in the advancing domain of machine learning for constitutive modeling.
Key facts
- Paper arXiv:2608.14063v1
- Benchmarks on Treloar's classic experimental data
- Compares four machine-learning-based hyperelasticity frameworks
- Evaluates fitting performance, computational cost, hyperparameter sensitivity, ease of implementation
- Discusses trade-offs between predictive accuracy and model complexity
- Assesses model complexity via number of material parameters and computational time
- Published on arXiv (announce type: new)
Entities
Institutions
- arXiv