hyper-surrogate 0.4.0

Data-driven surrogates for hyperelastic constitutive models in finite element analysis

hyper-surrogate

Define hyperelastic materials in Python, deploy them in your finite element solver.

• Github repository: https://github.com/jpsferreira/hyper-surrogate/
• Documentation: https://jpsferreira.github.io/hyper-surrogate/

Table of Contents

• Why hyper-surrogate?
• Architecture
• Features
• Quick Start
• Installation
• Examples
• Contributing

Why hyper-surrogate?

Finite element solvers need constitutive models. Writing and maintaining those models — especially user-material subroutines for Abaqus, FEAP and similar Fortran-based solvers — is repetitive: kinematics, second Piola–Kirchhoff stress, consistent tangent, Jaumann correction, reference-configuration freedom, all in Fortran 77 or 90.

hyper-surrogate lets you define your material once in Python and emit a self-contained Fortran 90 user subroutine deployable in any Fortran-based FE solver. You choose how the material is defined:

1. Built-in symbolic SEF — one of ten classical isotropic and anisotropic models (Neo-Hooke, Mooney-Rivlin, Yeoh, Holzapfel-Ogden, Gasser-Ogden-Holzapfel, …).
2. Custom symbolic SEF — subclass Material with a sef property in SymPy and the framework derives stress and tangent automatically.
3. Data-driven surrogate — train an MLP, ICNN, or polyconvex ICNN on energy and stress samples from any ground-truth source and emit a hybrid UMAT whose energy is the trained network and whose stress and tangent are computed analytically in Fortran.

All three paths produce the same artefact: a .f90 with Cauchy stress, analytical consistent tangent, and stress-freedom at the reference configuration enforced exactly at emission time. No Python runtime is required at solve time.

Architecture

Material ─> DeformationGenerator ─> Dataset ─> MLP / ICNN ─> HybridUMATEmitter ─> .f90
   │                                              │
   │         (symbolic SEF)                        │  (trained NN weights)
   │                                              │
   └──── UMATHandler ─────────────────────────────┘
         (analytical Fortran via SymPy CSE)

Features

• Three entry points, one output — built-in SEF, custom SymPy SEF, or trained surrogate, all emit a structurally identical Fortran UMAT.
• Symbolic mechanics — automatic PK2 stress and stiffness tensor derivation via SymPy; analytical Cauchy push-forward and Jaumann correction emitted with common-subexpression elimination.
• NN architectures — MLP, Input-Convex NN, and polyconvex ICNN, trained with an energy-plus-stress dual loss that backpropagates the gradient target through autograd.
• Hybrid UMAT — neural-network strain energy with closed-form analytical kinematics, stress, and consistent tangent in Fortran; layer-additive Hessian backpropagation verified against autograd at $10^{-5}$ tolerance.
• Stress freedom at reference — enforced exactly at emission time by subtracting a compile-time offset; no surprise residual stress at $\mathbf{C} = \mathbf{I}$.
• Anisotropic support — fiber pseudo-invariants $I_4, I_5$ for one or two fiber families with arbitrary orientations.

Quick Start

Three short pipelines, mirroring the three ways to define a material:

Path A — built-in symbolic SEF

from hyper_surrogate import NeoHooke, UMATHandler

UMATHandler(NeoHooke({"C10": 0.5, "KBULK": 1000.0})).generate("neohooke.f90")

Path B — custom user-defined SEF

import sympy as sp
from hyper_surrogate import Material, UMATHandler

class MyOgdenLike(Material):
    DEFAULT_PARAMS = {"mu": 1.0, "alpha": 2.0, "KBULK": 1000.0}

    def __init__(self, parameters=None):
        super().__init__({**self.DEFAULT_PARAMS, **(parameters or {})})

    @property
    def sef(self):
        h = self._handler
        mu, alpha, K = (self._symbols[k] for k in ("mu", "alpha", "KBULK"))
        return ((mu / alpha) * (h.isochoric_invariant1 ** (alpha / 2) - 3)
                + 0.5 * K * (sp.sqrt(h.invariant3) - 1) ** 2)

UMATHandler(MyOgdenLike()).generate("mysef.f90")

Path C — data-driven surrogate

from hyper_surrogate import (
    NeoHooke, MLP, Trainer, EnergyStressLoss,
    create_datasets, extract_weights, HybridUMATEmitter,
)

material = NeoHooke({"C10": 0.5, "KBULK": 1000.0})
train_ds, val_ds, in_norm, energy_norm = create_datasets(
    material, n_samples=4000, input_type="invariants",
    target_type="energy", deformation_mode="combined_compressible",
)
model = MLP(input_dim=3, output_dim=1, hidden_dims=[64, 64, 64], activation="softplus")
result = Trainer(model, train_ds, val_ds,
                 loss_fn=EnergyStressLoss(alpha=1.0, beta=1.0),
                 max_epochs=2000, patience=200).fit()
exported = extract_weights(result.model, in_norm, energy_norm)
HybridUMATEmitter(exported).write("neohooke_hybrid.f90")

See the documentation for the full tutorial set, including custom materials, anisotropic models, and the export-path decision tree.

Installation

pip install hyper-surrogate        # core (NumPy, SymPy)
pip install hyper-surrogate[ml]    # with PyTorch for ML surrogates

From source with uv:

git clone https://github.com/jpsferreira/hyper-surrogate.git
cd hyper-surrogate
uv sync --all-groups --extra ml

Examples

Runnable scripts are in the examples/ directory:

Script	Path	Description
analytical_umat.py	A	Built-in NeoHooke → analytical UMAT via UMATHandler
custom_sef.py	B	Custom Ogden-like SEF → analytical UMAT
export_hybrid_umat.py	C	End-to-end train + HybridUMATEmitter export
train_neohooke_sef.py	C	Train MLP on NeoHooke SEF with hybrid inference
train_icnn_energy.py	C	Train ICNN with EnergyStressLoss
train_polyconvex.py	C	Train PolyconvexICNN with per-invariant convexity
train_holzapfel_ogden.py	C	Anisotropic fiber-reinforced training pipeline

Run any example with:

uv run python examples/<script>.py

Contributing

Contributions are welcome! See CONTRIBUTING.md for setup instructions and guidelines.