Learning shared hyperelastic constitutive representation from heterogeneous full-field experiments across multiple materials
Abstract
A material-conditioned Physics-Augmented Neural Network is introduced for unsupervised multimaterial hyperelastic constitutive learning from experimental full-field data. The complete database comprises 30 experiments quantified by image correlation on various materials such as elastomers, cork, foam, and architected metamaterials. The database includes several specimen geometries, loading paths, mesh resolutions, and measurement modalities. The model consists of a shared input-convex backbone that maps strain invariants to convex energy modes. Material dependence is introduced only through non-negative modal amplitudes, yielding a compact constitutive embedding while preserving a common admissible backbone. Among the investigated embedding dimensions, the four-mode representation provides the best post-calibration performance, which corresponds to a reduction of 66% relative to independently calibrated Neo-Hookean models, while remaining less accurate than independently trained single-dataset PANNs. The learned amplitudes exhibit family-level organization in exploratory low-dimensional projections. The frozen backbone further enables reduced calibration of experimental instances withheld from backbone training, using only four non-negative coefficients. These results show that heterogeneous full-field experiments can be used collectively to learn a reusable, physically constrained constitutive basis, rather than being treated only as independent identification problems.
Related
- Data-Driven Discovery of a Hyperelastic Strain-Energy Function for Veins Using Constitutive Artificial Neural Networks
- Identification of constitutive materials of bi-layer soft tissues from multimodal indentations
- Data-Driven Discovery of Composition-Dependent Constitutive Models for Hyperelasticity and Viscoelasticity of Digital Materials
- A new perspective in linear Cauchy Elasticity: variational minimum principles for statics, dynamics, and heterogeneous materials
- Creep failure in heterogeneous materials from the barrier landscape
BibTeX
@article{mabileau2026learning,
title = {{Learning shared hyperelastic constitutive representation from heterogeneous full-field experiments across multiple materials}},
author = {Lucas Mabileau and Maxence Barberet-Pinto and Thomas Ferchal and Margot Leclercq and Jan Neggers and Remi Legroux and Pierre Latil and Malo Valmalle and Benjamin Smaniotto and François Hild and Emmanuel Baranger and Clément Jailin},
journal = {Computer Methods in Applied Mechanics and Engineering},
year = {2026},
month = oct,
doi = {10.1016/j.cma.2026.119498},
url = {https://hal.science/hal-05709206v2},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Learning shared hyperelastic constitutive representation from heterogeneous full-field experiments across multiple materials
AU - Lucas Mabileau
AU - Maxence Barberet-Pinto
AU - Thomas Ferchal
AU - Margot Leclercq
AU - Jan Neggers
AU - Remi Legroux
AU - Pierre Latil
AU - Malo Valmalle
AU - Benjamin Smaniotto
AU - François Hild
AU - Emmanuel Baranger
AU - Clément Jailin
PY - 2026
DA - 2026/10/30
JO - Computer Methods in Applied Mechanics and Engineering
DO - 10.1016/j.cma.2026.119498
UR - https://hal.science/hal-05709206v2
AB - A material-conditioned Physics-Augmented Neural Network is introduced for unsupervised multimaterial hyperelastic constitutive learning from experimental full-field data. The complete database comprises 30 experiments quantified by image correlation on various materials such as elastomers, cork, foam, and architected metamaterials. The database includes several specimen geometries, loading paths, mesh resolutions, and measurement modalities. The model consists of a shared input-convex backbone that maps strain invariants to convex energy modes. Material dependence is introduced only through non-negative modal amplitudes, yielding a compact constitutive embedding while preserving a common admissible backbone. Among the investigated embedding dimensions, the four-mode representation provides the best post-calibration performance, which corresponds to a reduction of 66% relative to independently calibrated Neo-Hookean models, while remaining less accurate than independently trained single-dataset PANNs. The learned amplitudes exhibit family-level organization in exploratory low-dimensional projections. The frozen backbone further enables reduced calibration of experimental instances withheld from backbone training, using only four non-negative coefficients. These results show that heterogeneous full-field experiments can be used collectively to learn a reusable, physically constrained constitutive basis, rather than being treated only as independent identification problems.
SN - 0045-7825
SN - 1879-2138
ER -