Port-Hamiltonian Neural Networks with Diffusive Fractional Memory and Tensor-Train Operators: Passivity-Preserving Surrogates for Digital Twins of Viscoelastic Structures
Neural surrogates of structures with fractional viscoelastic damping must carry the memory of the Caputo derivative and, over long horizons, must not create energy; penalty-based training guarantees neither. We propose TT-fpHNN, a port-Hamiltonian network in which the diffusive representation of the Caputo derivative with positive quadrature weights yields a finite port-Hamiltonian system for any number of memory nodes. Tail corrections with explicit bounds reduce the kernel error of a 24-node rule from 6.4% to 0.09%, a discrete-gradient integrator keeps the energy balance exact, and the learned operators are positive semidefinite tensor-train factors. On synthetic trusses, structured rollouts kept a per-step balance residual below $10^{-12}$, whereas a sum-of-exponentials scheme of equal memory and accuracy accumulated a 1.7% energy imbalance. On 450 degrees of freedom, the rank-8 model reproduced 200 periods with 0.66% error and 11,106 parameters; an unconstrained model diverged, a model without memory reached 3.4%, and a regularised dense model needed 36.5 times more parameters. On 2160 degrees of freedom the model stayed passive over 10,000 steps with 3.3% error. Rank 16 over-fitted to a nearly singular stiffness. The fractional order was recovered within $\pm0.05$ at 1% noise with two-decade excitation and biased at 5%. Single-precision materialisation at condition number $7.7\times10^{7}$ destroyed definiteness.
Authors
- José García (ORCID: https://orcid.org/0000-0003-3126-8352)
- Luis Rojas (ORCID: https://orcid.org/0009-0005-8935-7595)
Publication Details
- Journal
- Preprints.org
- Published
- 2026-11-17
- DOI
- https://doi.org/10.20944/preprints202611.1680.v1
- Primary Topic
- Model Reduction and Neural Networks
- Type
- preprint