TorchFEA — A PyTorch Finite-Element Analysis Framework

Developed by Zenan Song

TorchFEA is a research-oriented finite-element framework built on PyTorch. It provides a single object model for parts, assemblies, materials, loads, constraints, solvers, and results, with automatic differentiation available through the same tensor-based formulation. Models can be authored from Python, imported from Abaqus-style .inp files, or built interactively in the Gmsh/OCC-based CAE interface.

Features

• Nonlinear three-dimensional solid mechanics with PyTorch tensors and float64 arithmetic
• Static implicit Newton–Raphson solving with line search and sparse Pardiso/Conjugate-Gradient linear algebra
• Dynamic implicit (Newmark-β) and dynamic explicit (central-difference) solvers
• Hyperelastic and Saint Venant–Kirchhoff/linear-elastic material models
• Pressure, body force, concentrated force, moment, springs, penalty degrees of freedom, self-contact, and body-to-body contact
• Reference points, node coupling, displacement boundaries, named node/element sets, and surface sets
• PyTorch automatic differentiation for Jacobians, sensitivities, and optimization workflows
• Native Gmsh OCC CAD histories with boxes, wedges, cylinders, spheres, cones, tori, STEP import, transforms, and Boolean operations
• Abaqus-style CAE preprocessing and postprocessing UI with BREP/mesh views, selection, contour plots, deformation-gradient, stress, strain, and result export
• Native .npz model/result serialization with model check codes for result validation

Installation

TorchFEA requires Python ≥ 3.12. A PyTorch float64 build is recommended; CUDA is optional and can be selected through PyTorch's normal device settings.

Install a released package:

pip install torchfea

Install the current checkout while developing:

pip install -e .

The GUI entry point is installed with the package:

torchfea-ui

When running directly from the source tree, use:

python -m torchfea.ui

Quick start: INP model and static solve

The following is the same construction pattern used by the pressure examples:

from pathlib import Path

import torch
import torchfea

torch.set_default_dtype(torch.float64)

inp = torchfea.FEA_INP()
inp.read_inp(Path("tests/models/C3D4.inp"))

fe = torchfea.from_inp(inp)
fe.solver = torchfea.solver.StaticImplicitSolver(tol_error=1e-8)

fe.assembly.add_load(
    torchfea.loads.Pressure(
        instance_name="final_model",
        surface_set="surface_1_All",
        pressure=0.06,
    ),
    name="pressure-1",
)
fe.assembly.add_boundary(
    torchfea.boundarys.Boundary_Condition(
        instance_name="final_model",
        set_nodes_name="surface_0_Bottom",
        indexDoF=[0, 1, 2],
    ),
    name="fixed-bottom",
)

result = fe.solve()
print(result.converged, result.GC.shape)

fe.save_model("model.npz")
result.save("result.npz")

The same Assembly API is used for reference points, coupling constraints, springs, contact, additional instances, materials, and boundary conditions. See docs/usage.md for complete examples.

CAD construction from Python

CAD is represented as a serializable feature history owned by FEAController.cad. The history is replayable and independent of the UI:

from torchfea import FEAController
from torchfea.cad import Box, Cylinder, Cut

fe = FEAController()
history = fe.cad.add_part("Part-1")

history.add(Box("Box-1", dx=20, dy=10, dz=5))
history.add(Cylinder("Cylinder-1", x=10, y=5, radius=2, dz=5))
history.add(Cut("Cut-1", object_id="Box-1", tool_id="Cylinder-1"))

# Generate a C3D4 Part with one global mesh size.
part = fe.cad.mesh_part("Part-1", mesh_size=1.5)

Every feature is an explicit serializable class with a discoverable Python signature. The available operations are Box, Wedge, Cylinder, Sphere, Cone, Torus, StepImport, Fuse, Cut, Intersect, Fragment, Translate, Rotate, Dilate, and Mirror. CADHistory also provides typed convenience methods such as add_box, cut, and rotate. After meshing, attach the returned native Part to an Assembly, assign materials and density to its element blocks, add an Instance, and define the analysis conditions through the normal model API. The CAE UI performs these same kernel operations and saves the history through FEAController.save_model.

CAE workflow

The graphical preprocessor and postprocessor are launched with torchfea-ui. The preprocessor supports:

1. Gmsh/OCC geometry histories and Boolean features
2. Part materials, density, node sets, element sets, and surfaces
3. Reference points, instances, boundaries, couplings, and loads
4. One project-wide mesh size and FE mesh generation
5. Part/assembly geometry and mesh views with BREP picking
6. Native model save, INP export, and STEP import/export

The postprocessor can inspect all instances or one instance, display displacement, deformation-gradient, Green–Lagrange strain, Cauchy stress, Mises values, volume ratio, and strain energy, and export .npz or .csv results. Both CAD and mesh views provide standard camera views and section clipping.

An .npz created from an INP model is treated as a discrete-mesh model: it retains the existing nodes, elements, sets, materials, loads, and constraints without inventing a CAD history or remeshing it. A CAD-created model retains its OCC history and can be rebuilt in the UI.

See docs/cae_ui.md for the complete CAE workflow.

Persistence and result validation

Use the controller serializer for a complete native model:

import torchfea

model = torchfea.load_model("model.npz")
result = model.solve()

FEAController.save_model stores the assembly, discrete parts, CAD histories, materials, sets, loads, constraints, boundaries, and solver settings in the native torchfea representation. StaticResult.save/load and DynamicResult.save/load store solver results separately. Model and result archives carry a compact model check code; the UI rejects a result whose code or degree-of-freedom layout does not match the loaded model.

Repository layout

src/torchfea/
├── controller.py       # model construction, solve, and serialization
├── cad/                # Gmsh OCC history and meshing
├── model/              # Assembly, Part, elements, materials, loads, BCs
├── solver/             # static and dynamic solvers/results
└── ui/                 # PySide6 + VTK CAE interface

docs/                   # usage, architecture, solver, loads, and CAE guides
tests/                  # element, pressure, contact, dynamics, and sensitivity cases

Useful references:

• docs/usage.md — Python model definition and solving
• docs/cae_ui.md — CAE pre/postprocessor workflow
• docs/structure.md — package architecture and data flow
• docs/solver.md — nonlinear and dynamic solver details
• docs/loads.md — load interface and examples

License and scope

TorchFEA is research software. For production or safety-critical engineering use, validate the formulation, material parameters, mesh convergence, and solver settings for the target problem before relying on the results.