New AI Technique Decodes Soft Material Physics From a Single Experiment
Chemistry

New AI Technique Decodes Soft Material Physics From a Single Experiment

Advanced neural networks can now learn complex hyperelastic material properties from heterogeneous tests without human guesswork.

By Anjali Singh
Published:
Email this Article

Soft materials such as rubber and biological tissue stretch in complex ways that defy simple description, but a new computational technique can automatically decode their mechanical behavior from a single complex experiment. Researchers have developed an advanced neural network framework that eliminates the need for tedious human guesswork when formulating material equations.

Predicting how soft structures deform under load usually requires an expert to manually select a mathematical formula and run multiple uniform stress tests. The new study, published in arXiv by Kian Abdolazizi, Kevin Linka, and Christian Cyron, bypasses this bottleneck by combining specialized machine learning models with advanced finite element simulations.

How Neural Networks Capture Material Physics

Traditional neural networks act as opaque black boxes that often violate basic physical laws, such as thermodynamic consistency and material symmetry. To overcome this limitation, the authors utilized Constitutive Kolmogorov–Arnold Networks, or CKANs. These architectures bake fundamental physical requirements directly into their structure by construction rather than learning them from data.

Instead of learning fixed weights on static activation functions, CKANs learn flexible univariate functions parameterized as B-splines along the network edges. This design preserves crucial mathematical properties like convexity and monotonicity, ensuring that the model remains stable during computer simulations of extreme deformation.

Translating Complex Networks into Simple Formulas

While neural networks offer high flexibility, engineers prefer clean, closed-form algebraic equations for practical design work. To bridge this gap, the research team applied two post-processing steps called pruning and symbolic regression. Pruning strips away redundant edges and nodes with negligible attribution scores, while symbolic regression replaces spline functions with standard mathematical terms from a predefined library.

The resulting expressions automatically reduce to classical models like the Neo-Hookean and Mooney–Rivlin strain energies depending on the training dataset. This workflow transforms black-box machine learning outputs into interpretable formulas that engineers can easily inspect, verify, and implement into commercial simulation software.

Testing the Approach on Rubber and Aortic Tissue

The team validated their framework using three distinct experimental datasets involving large deformations and complex geometries. In the first test involving perforated natural rubber membranes, full-field displacement data captured via digital image correlation supervised the network training. The trained CKANs successfully reproduced measured displacement fields with median errors between one and two percent.

In a second experiment involving a silicone plate with a cut ligament, the network learned successfully from a single one-dimensional stretch profile along an internal path. Finally, the researchers tested an anisotropic human aortic tissue sample inflated under fluid pressure. By incorporating fiber reinforcement into the network architecture, the model accurately predicted the three-dimensional bulging behavior of the arterial wall.

Limitations and Future Directions

Despite these successes, the authors note that the quality of the discovered material law depends heavily on the coverage of the deformation space rather than the sheer volume of data. If an experiment fails to probe certain stretching states, the model may extrapolate poorly outside its training domain. Future investigations may explore adaptive grid techniques that adjust during training to better capture incoming activation ranges.

The study was posted to the arXiv preprint server on November 19, 2026.

Fact Checked

This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.

Last reviewed on .

Article history

  • Latest version

Cite this page:

Singh, Anjali. “New AI Technique Decodes Soft Material Physics From a Single Experiment.” BioScience. BioScience ISSN 2521-5760, 20 November 2026. <https://www.bioscience.com.pk/en/subject/chemistry/constitutive-kolmogorov-arnold-networks-in-finite-element-model-updating-closed-form-hyperelastic-strain-energies-from-heterogeneous-experiments>. Singh, A. (2026, November 20). “New AI Technique Decodes Soft Material Physics From a Single Experiment.” BioScience. ISSN 2521-5760. Retrieved November 20, 2026 from https://www.bioscience.com.pk/en/subject/chemistry/constitutive-kolmogorov-arnold-networks-in-finite-element-model-updating-closed-form-hyperelastic-strain-energies-from-heterogeneous-experiments Singh, Anjali. “New AI Technique Decodes Soft Material Physics From a Single Experiment.” BioScience. ISSN 2521-5760. https://www.bioscience.com.pk/en/subject/chemistry/constitutive-kolmogorov-arnold-networks-in-finite-element-model-updating-closed-form-hyperelastic-strain-energies-from-heterogeneous-experiments (accessed November 20, 2026).
End of the article