🤖 AI

Benchmarking data-driven material models on the classic Treloar dataset

This paper benchmarks six popular data-driven constitutive modeling frameworks on the classic Treloar dataset to compare their performance, computational cost, and complexity, ultimately providing practical guidance on their respective strengths and limitations rather than declaring a single superior method.

Original authors: Hagen Holthusen, Moritz Flaschel, Denisa Martonová, Ellen Kuhl

Published 2026-08-17

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Hagen Holthusen, Moritz Flaschel, Denisa Martonová, Ellen Kuhl

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to recreate the perfect flavor of a secret sauce. You have a list of ingredients (the data) and a goal: to write down a recipe (a mathematical model) that tastes exactly like the original. In the world of engineering, this "sauce" is how materials like rubber, skin, or tires stretch and squish. For decades, scientists have tried to guess the recipe by picking from a limited menu of pre-written formulas. But now, a new wave of technology called machine learning is stepping into the kitchen. Instead of guessing, these smart algorithms can taste the data and try to invent the recipe themselves. The big question is: if you let a computer cook up a new recipe, will it actually work? Will it be fast enough to use in real-time simulations, or will it be a slow, complicated mess that crashes the computer?

This paper puts six different "AI chefs" to the test in a high-stakes cooking competition. The judges used a classic, famous dataset of rubber stretching experiments (known as the Treloar dataset) as their taste test. They asked these six methods to learn the recipe for rubber based on how it stretches in three different ways: pulling it like a rubber band, stretching it like a balloon, and shearing it like a deck of cards. The goal wasn't just to see who could match the data best, but to see who could do it with the simplest recipe, the least amount of computer time, and the most consistent results.

The results were a bit like a reality TV show where no single contestant won every challenge. The paper found that all six methods were incredibly good at mimicking the rubber's behavior, almost perfectly matching the experimental data. However, they each had very different strengths and weaknesses. One method, called Adaptive Material Fingerprinting (AMF), was the most accurate chef, creating a model that matched the data with a score of 0.9996 out of 1.0. Another method, Generalized Invariant-based CANN (GI-CANN), was the "smartest" chef, achieving nearly the same high accuracy but with a much simpler recipe that only required eight adjustable numbers (parameters).

On the other end of the spectrum, some methods were incredibly fast but less precise. Material Fingerprinting (MF) was the speed demon, figuring out the recipe in just 0.302 milliseconds—faster than you could blink. However, its recipe was a bit rigid, and it struggled a little more when the rubber was stretched in a specific way called "equibiaxial tension." Meanwhile, the neural network-based chefs (like PANN) were flexible but required a massive amount of computer power to train and ended up with very complicated recipes containing hundreds of parameters, without necessarily tasting any better than the simpler ones.

The paper explicitly rules out the idea that there is one single "best" method for every situation. Instead, it suggests that the right choice depends on what you value most. If you need the absolute highest accuracy and don't mind a complex model, AMF is the way to go. If you need a model that is both highly accurate and simple enough to run quickly in a simulation, GI-CANN offers the best balance. If you need to identify materials instantly and can afford to do the heavy lifting beforehand, the database-driven MF method is unbeatable. Ultimately, the authors show that while machine learning can successfully discover how materials work, the "best" tool depends entirely on the specific job you need it to do.

Technical Summary: Benchmarking Data-Driven Material Models on the Classic Treloar Dataset

Problem Statement
The emergence of machine learning (ML) and automated model discovery has transformed constitutive modeling, offering alternatives to traditional phenomenological and micromechanical approaches. These data-driven methods can learn strain energy density functions directly from experimental data, balancing physical structure, flexibility, sparsity, and interpretability. However, the growing number of available frameworks has created a need for systematic comparison. While classical benchmarks (e.g., Steinmann et al., 2012) evaluated phenomenological models against the classic Treloar rubber dataset, a renewed benchmark is required to assess how modern data-driven approaches compare in terms of predictive accuracy, computational cost, hyperparameter sensitivity, and model complexity.

Methodology
The authors benchmark six representative data-driven constitutive modeling approaches under identical conditions using the classical Treloar natural rubber dataset (Treloar, 1944). The dataset comprises stress-stretch measurements from three homogeneous deformation modes: uniaxial tension (UT), equibiaxial tension (ET), and pure shear (PS). The UT and ET data serve as the training/calibration set, while the PS data are reserved for independent validation to assess generalization.

The six methods evaluated are:

Efficient Unsupervised Constitutive Law Identification & Discovery (EUCLID): Uses sparse regression (L1-regularization) on a library of candidate modeling terms (e.g., generalized Mooney-Rivlin) to identify a parsimonious set of constitutive terms.
Material Fingerprinting (MF): A database-driven approach that replaces continuous optimization with pattern recognition. It compares normalized experimental stress responses against a pre-computed database of fingerprints from various constitutive models.
Adaptive Material Fingerprinting (AMF): An extension of MF that incrementally constructs the constitutive model by selecting individual modeling features from a database, rather than selecting a single pre-defined model.
Physics-Augmented Neural Networks (PANNs): Utilizes Input-Convex Neural Networks (ICNNs) to ensure polyconvexity and physical consistency, taking principal invariants as inputs.
Constitutive Artificial Neural Networks (CANNs): Employs a structured neural network architecture combining physically motivated invariant-based features with trainable activation functions and non-negative weights.
Generalized Invariant-based CANNs (GI-CANNs): Extends CANNs by treating the invariant representation itself as trainable, allowing the discovery of generalized invariants (powers of principal stretches) alongside the energy function.

Evaluation Metrics
The study employs a multi-faceted evaluation framework:

Predictive Accuracy: Measured via Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and the coefficient of determination ( $R^2$ ) on both training and validation (pure shear) datasets.
Model Complexity: Quantified by the effective number of non-zero parameters ( $\|\theta\|_0$ ) and the computational time required to evaluate the strain energy density ( $W$ ), its gradient (stress, $P$ ), and Hessian (tangent, $H$ ) during finite element simulations.
Computational Cost: Distinguished between the time required for model identification (training/discovery) and the runtime for constitutive evaluation.
Robustness: Assessed by analyzing variability across multiple random initializations (for neural networks) or hyperparameter configurations (for adaptive methods).

Key Results

Predictive Performance: All six methods successfully reproduce the benchmark data with high accuracy. AMF achieved the highest validation accuracy ( $R^2 = 0.9996$ ), followed closely by GI-CANN ( $R^2 = 0.9993$ ). EUCLID and GI-CANN demonstrated particularly strong generalization to the pure shear validation set. PANN showed the highest variability across initializations, particularly for equibiaxial tension.
Accuracy-Complexity Trade-off:
- MF produced the sparsest model (2 parameters) with the lowest evaluation cost but showed larger deviations in equibiaxial tension.
- EUCLID offered a favorable compromise with only 4 parameters and very low identification cost, achieving high accuracy.
- GI-CANN provided a strong balance, achieving top-tier accuracy with only 8 parameters.
- PANN required significantly more parameters (371) without a corresponding gain in validation accuracy compared to simpler models.
Computational Cost:
- Identification: EUCLID and MF were orders of magnitude faster (sub-millisecond core times) than neural network methods (hundreds of milliseconds to seconds), as the latter require iterative gradient-based optimization.
- Evaluation: MF was the fastest for evaluating $W$ , $P$ , and $H$ . Interestingly, the number of parameters did not strictly dictate evaluation cost; for instance, PANN's evaluation time was comparable to other neural network models despite having more parameters.
Robustness: Methods relying on sparse regression or database matching (EUCLID, MF) produced deterministic results. Neural network methods (PANN, CANN, GI-CANN) exhibited sensitivity to initialization, with GI-CANN showing a more robust optimization landscape than PANN.

Significance and Claims
The paper does not identify a single "winner" among the six methods. Instead, it highlights the distinct strengths and limitations of each approach to provide practical guidance for selection:

For applications prioritizing computational efficiency and sparsity: EUCLID and MF are preferred due to their low identification costs and compact representations.
For applications prioritizing predictive accuracy and flexibility: AMF and GI-CANN offer superior performance, with AMF achieving the highest validation accuracy and GI-CANN providing a particularly favorable balance of high accuracy and moderate complexity.
For robustness: GI-CANN appears more robust to initialization than PANN, suggesting that generalized invariant representations may offer a more stable optimization landscape.

The authors conclude that the choice of method depends on the specific requirements of the application, such as whether predictive accuracy, model sparsity, robustness, or computational efficiency is the highest priority. The study emphasizes that good agreement in training modes does not guarantee identical performance in unseen deformation modes, underscoring the importance of independent validation. All source code, training scripts, and data used in the study are publicly available to facilitate reproducibility.

More like this