Skip to main content
Back to timeline
arXivSource publication:

A few specimens can define a whole family of hyperelastic laws: hierarchical Bayesian PANNs separate specimen spread from noise

Synopsis

A hierarchical Bayesian physics-augmented neural network (PANN) learns the constitutive laws of several specimens of one material as a population: all specimens share polyconvex features learned by the network, each specimen is described only by its own amplitude and feature shares, and the population mean and covariance and the measurement noise are inferred in one posterior. 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 its bands for untested curves are too narrow.

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

Interpretation

The constitutive law differs between specimens of one material, and often only a few specimens can be tested; the new model treats each specimen's law as a draw from a population instead of pooling all measurements into one mean law. Earlier pooled fits attributed the differences between specimens to measurement noise or to uncertainty about a mean law, and neither describes a new specimen. On synthetic populations with a known ground truth, the population learned from few specimens beats the Gaussian CANN, the Bayesian CANN and a diffusion-based method on energy distance, coverage and band width, and the Gaussian CANN catches up on coverage and width only at twelve specimens.

Each specimen is described by two interpretable quantities: a positive amplitude that sets its stress level and a set of positive shares summing to one that set how much each feature contributes to the shape of its law. All specimens share the polyconvex features learned by the hidden layers, so a partly tested specimen borrows them from the others. Features, specimen coordinates, population and a separate measurement noise are inferred jointly in one posterior by variational inference; on twelve healthy and aneurysmal aortas under leave-one-out, ten specimens are assigned to their true cohort with a probability of at least 0.9, two are undecided, and none is assigned to the wrong cohort with confidence.

Given a short uniaxial test of a new specimen, the population narrows the prediction from the whole population to that specimen and extrapolates to untested load cases. Without a population, a single-specimen fit follows the data only within the tested range and continues by its prior beyond it, missing the specimen's upturn. On synthetic populations, when the uniaxial test ends at 1.25 or 1.5, the band covers about 0.8 to 0.9 of the true equibiaxial and plane-strain curves, close to a reference given the correct form of the law; with the complete uniaxial curve the coverage falls to about 0.6, showing overconfidence.

Perspective

The demonstrated setting is homogeneous uniaxial, equibiaxial and plane-strain loading on synthetic hyperelastic laws, Treloar's vulcanised rubber, and biaxial data from healthy and aneurysmal human abdominal aortas; the aorta data were digitised from published figures and curves ending early were completed with synthetic points.

The population is assumed Gaussian, and the authors note that skewed or multimodal populations require richer families such as diffusion models, which in turn need more specimens; with only three specimens the model is overconfident about the width of the population, and a long, precise test locates a new specimen more sharply than the data justify, lowering coverage in untested load cases; one general prior was kept unchanged in all experiments, and the authors state that it should be chosen according to available prior knowledge of the material.