Hierarchical Bayesian physics-augmented neural networks for populations of hyperelastic laws: specimen variability and uncertainty from scarce, noisy data

Vincent C. Scholz, Phaedon-Stelios KoutsourelakisarXiv 2026 · arXiv

This paper presents a significant step forward in constitutive modeling by combining the flexibility of neural networks with the rigor of hierarchical Bayesian statistics. The authors introduce a Hierarchical Bayesian Physics-Augmented Neural Network (PANN) designed to discover the mechanical laws of materials—specifically hyperelastic tissues—while accounting for the inherent variability between different specimens.

TL;DR

The authors develop a hierarchical Bayesian framework for physics-augmented neural networks that separates measurement noise from inherent material variability in hyperelastic populations. This allows for the discovery of physically admissible material laws from scarce, noisy data, enabling accurate predictions for new specimens and reliable classification of tissue cohorts (e.g., healthy vs. diseased).

What Problem Does This Paper Address?

In biomechanics, the "law" that governs how a tissue deforms under load (its constitutive law) varies from person to person. Researchers often face two major hurdles:

  1. Data Scarcity and Noise: Testing biological tissue is difficult, resulting in few samples and noisy measurements.
  2. The "Pooling Trap": Standard approaches often "pool" data from multiple specimens, treating them as repeated measurements of a single "mean" law. As shown in Figure 1, this mistake leads to a model that confuses specimen variability with measurement noise, resulting in overconfident mean predictions that fail to describe any individual new specimen.
Figure 1: The pooling trap (Section 1.2), illustrated on a toy problem with six specimens. (a) A pooled fit has no place for the spread of the specimens, so it attributes the spread to the uncertainty of the mean law and to a larger measurement noise. (b) A population model (Section 3.2) describes the spread by the population from which the specimens are drawn. The two bands in (a) together can look as plausible as the band in (b), but they describe something different. The band of the mean law shrinks as more measurements come in, whereas the population band keeps the spread of the specimens. A draw from the pooled fit is the mean curve with independent noise at every point, whereas a draw from the population is the curve of a specimen.

Existing methods either rely on pre-defined functional forms (which might be wrong) or black-box neural networks (which might violate physics). This paper seeks a middle ground: a model that learns the form from data but remains physically valid and statistically sound.

Methods

The core of the approach is a Hierarchical Bayesian PANN. It builds on the Physics-Augmented Neural Network (PANN) architecture, which uses Input Convex Neural Networks (ICNNs) to ensure that the learned energy density is polyconvex—a key requirement for physical admissibility.

The "Hierarchical" part is what handles the population. As illustrated in Figure 2, the model structure is split:

  • Shared Features: The hidden layers of the neural network are shared across all specimens. These represent the fundamental "building blocks" of the material's behavior learned from the entire group.
  • Specimen-Specific Weights: Each individual specimen $k$ is characterized by its own set of weights ($w_k$). These weights determine the amplitude (how stiff it is) and the shares (the specific mix of features that define its shape).
  • Population Prior: These specimen weights are treated as draws from a global Gaussian population.
Figure 2: Structure of the hierarchical Bayesian PANN, which separates what all specimens share from what differs between them. The hidden layers produce N polyconvex features, shared by all specimens. Specimen k weights the normalised features (9) by an amplitude exp(ak), which sets its stress level, and by shares π(zk) that sum to one, which set the shape of its response. The specimen parameters wk = (ak, zk) are draws from a Gaussian population (15). Since every draw is a positive combination of polyconvex features, every specimen of the population is polyconvex.

The authors use Variational Inference (VI) to jointly infer the features, the individual specimen parameters, and the population statistics (mean and covariance) in a single posterior distribution.

Datasets

The model was evaluated on a mix of synthetic and real-world data:

  • Synthetic Populations: Created using "S-shape" and "upturn" hyperelastic laws to test the model against a known ground truth at varying noise levels (1% to 25%).
  • Treloar’s Rubber: A classic benchmark dataset for vulcanized rubber.
  • Human Aortas: Data from 12 human abdominal aortas (6 healthy, 6 aneurysmal). This dataset is particularly challenging due to high variability and the need for clinical cohort assignment.

Key Insights and Ideas

  • Admissibility by Construction: Because the underlying PANN architecture enforces polyconvexity, every specimen "drawn" from the learned population is guaranteed to be physically plausible.
  • Separating Noise from Spread: By explicitly modeling measurement noise separately from the population covariance, the model avoids the "pooling trap." It understands that a specimen might deviate from the mean consistently across its entire stress-stretch curve, rather than just having random "jitter" at each point.
  • Feature Sharing: A partially tested specimen can "borrow" features learned from more extensively tested specimens in the same population, significantly improving predictions in untested load cases.

Key Results

The authors compared their work against several state-of-the-art methods, as summarized in the table below:

Table 1: Comparison of the proposed approach with existing methods across model class, calibration, and use requirements.
  • Robustness to Noise: On synthetic data with 25% noise, the Bayesian PANN provided smooth, plausible curves, while deterministic versions overfitted the noise.
  • Population Recovery: With as few as 6 specimens, the model recovered the true population distribution more accurately than competitors like Gaussian CANNs or diffusion-based models.
  • Clinical Utility: On the aorta data, the model correctly assigned 10 out of 12 specimens to their true cohort (healthy vs. aneurysmal) with high confidence (>90%). Crucially, it left the ambiguous cases as "undecided" rather than making confident wrong predictions.
  • Cross-Load Prediction: The model could predict the axial response of an aorta specimen using only its circumferential test data more accurately than a model fitted to that specimen alone.

Limitations and Future Work

  • Overconfidence with Tiny Samples: When the number of specimens is extremely low (e.g., $K=3$), the model tends to underestimate the true width of the population uncertainty.
  • Gaussian Latent Space: The assumption that specimen parameters follow a Gaussian distribution might be too simple for highly complex or multimodal populations. The authors suggest exploring mixture models or more flexible distributions in future work.
  • Prior Sensitivity: While the general priors worked well across all experiments, the authors note that incorporating specific domain knowledge into the priors could further improve performance for specialized materials.

Summary updated

Abstract

The hyperelastic law of a material such as soft tissue differs between specimens, and often only a few specimens can be tested. Physics-augmented neural networks (PANNs) learn such laws from data without a prescribed functional form while keeping them physically admissible, and Bayesian versions of related models quantify the uncertainty of the law of one specimen or of a pooled group of specimens. A pooled fit, however, attributes the differences between specimens to measurement noise or to uncertainty about a mean law, and neither describes a new specimen. We first develop a Bayesian PANN for one specimen, whose priors keep the learned law plausible for few, noisy data. We then extend it to a hierarchical Bayesian PANN, a nonlinear mixed-effects model in which all specimens share learned polyconvex features and each specimen is a draw from a population of laws. Features, specimens, and population are inferred jointly by variational inference, with a separate model of the measurement noise, so that the variability between specimens is not mistaken for noise, the population carries its own uncertainty, and every specimen drawn from it is polyconvex. The population predicts the response of new specimens with credible intervals, without data or from a partial test, and assigns them to cohorts with quantified uncertainty, all without a prescribed form of the law. On synthetic populations, it recovers the population from few specimens and predicts new specimens more reliably than three published methods. On Treloar's rubber and on healthy and aneurysmal aortas, it reaches the accuracy of a deterministic PANN and assigns no aorta specimen to the wrong cohort with confidence, although it underestimates the uncertainty of untested curves.
BibTeX
@misc{scholz2026hierarchical,
  author = {Vincent C. Scholz and Phaedon-Stelios Koutsourelakis},
  title = {{Hierarchical Bayesian physics-augmented neural networks for populations of hyperelastic laws: specimen variability and uncertainty from scarce, noisy data}},
  year = {2026},
  eprint = {2611.17892},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17892},
}
Figure