2026.07.22 v0.2.0

TensorMesh 0.2.0 is the mixed-assembly release: multi-field block systems become first-class, three new operator families land (Bloch-Floquet, distributed, open-domain wave), and the documentation — including the Chinese translation — is fully caught up.

Highlights

Mixed multi-field assembly (MixedElementAssembler) — #38

• Declare fields once (Field(trial="u", test="v", order=2, components=2)), write the coupled weak form as one scalar integrand, get one block SparseMatrix. The unchanged Condenser applies on top.
• assembler.layout (BlockLayout) owns the block-DOF vocabulary: boundary masks, split/cat, prolong/restrict, coordinate predicates.
• Generalized order pairs: field order is decoupled from mesh order via topological Lagrange DOF maps — Taylor-Hood P2-P1 runs directly on linear gmsh imports (no re-meshing), P3-P2 pairs and 3D P2-on-P1 tets included, all verified to machine precision.
• Space-aware load vectors via assemble_vector (+ the field_data channel for Picard / time stepping on generalized fields).
• The entire fluid example gallery now runs on stabilization-free Taylor-Hood: Stokes convergence study (H1/L2 rates 2.0/2.0), lid-driven cavity 2D/3D, transient cylinder flow, flow past obstacles, three-field Rayleigh-Bénard, Taylor-Green (velocity L2 rate ≈ 3, up from 2 with stabilized P1-P1).

Bloch-Floquet periodic boundary conditions (BlochReducer) — #36

Ties opposite unit-cell faces with the Floquet phase and reduces assembled operators to the master DOFs — the building block of band-structure computations. Ships with three COMSOL-validated 2D phononic-crystal examples (square/triangular lattices: mean band error 0.08% / 0.51%; slab transmission mean |ΔT| 0.007). Contributed by Nicolas Cayuela.

Distributed multi-GPU FEM (tensormesh.distributed)

DistributedMesh partitioning (RCB / spectral / METIS), the @distributed decorator that turns any existing assembler into its distributed counterpart, DSparseMatrix (a torch_sla.DSparseTensor subclass), and Condenser dispatch — end-to-end multi-GPU Poisson under torchrun with NCCL. Contributed by walkerchi.

Open-domain wave operators (PML & ports) — #39

• FacetBilinearAssembler — boundary bilinear forms scattered to a SparseMatrix (Robin / impedance / radiation).
• cartesian_pml + AnisotropicLaplaceElementAssembler / ScaledMassElementAssembler — stretched-coordinate PML with complex tensor coefficients.
• robin_operator / port_source — first-order absorbing / plane-wave-port boundary matrix and load.
• Four validated examples: port-driven Helmholtz resonator, PML-framed optical microdisk (whispering-gallery mode), rectangular waveguide modes, cantilever vibration modes + FRF. Contributed by Nicolas Cayuela.

Other improvements

• Drucker-Prager plasticity promoted to public API (DruckerPragerPlasticity, FrictionalMaterial presets) with geomechanics examples — kilickursat; Allen-Cahn convex-concave splitting example — MariusZeinhofer.
• Complex-valued assembly path (complex dtypes end-to-end through assemble → condense → solve) with a validated Helmholtz example.
• torch-sla alignment: requires torch-sla ≥ 0.3.0 (six verified solver backends); the [cupy] extra is replaced by [pyamg] (pip install "tensormesh-fem[gpu]" now pulls cuDSS + PyAMG). The old cuDSS SPD-only caveat is gone — general non-symmetric systems solve at machine precision.
• Fix: MeshGen.remove_sphere no longer hangs on spheres smaller than the mesh size (#40 → #47).

Documentation

New user-guide chapters (Mixed Assembly, Distributed FEM), new gallery chapters (Phononic Crystals, Open-Domain Wave, Modal Analysis, Taylor-Hood Stokes), a rewritten fluid chapter and solver-backend matrix, refreshed figures throughout — bilingual, with the zh_CN catalogs fully translated: https://docs.tensor-mesh.com

Breaking / migration notes

• tensormesh-fem[cupy] no longer exists (upstream torch-sla dropped its cupy backend); use the default CPU stack, [cudss], [pyamg], or [gpu].
• torch-sla < 0.3.0 is no longer supported.

Full changelog: v0.1.1...v0.2.0

2025.05.23v0.1.1

Fix: import tensormesh no longer plants a broken gmsh LazyLoader stub in sys.modules. Resolves '_LazyModule' object has no attribute 'initialize' on Python 3.12 + gmsh 4.15 (e.g. Colab). Mesh generation and python -m tensormesh.verify_install now work.

2025.05.23v0.1.0

🎉 First public release of TensorMesh!

TensorMesh is a fast, differentiable, JIT-free finite element method (FEM) library built natively on PyTorch. Write custom weak forms in pure Python and get tensorized assembly, sparse linear algebra, boundary conditions, and time integration — on CPU or GPU, with full autograd.

Install

pip install tensormesh-fem          # CPU
pip install "tensormesh-fem[gpu]"   # + CUDA sparse solvers (CuPy + cuDSS)

The import name stays import tensormesh. Verify with python -m tensormesh.verify_install.

Highlights

• GPU-native & differentiable — move an entire FEM workflow to GPU in one line; autograd flows through assembly and solve.
• High-performance tensorized assembly — a fully tensorized Map-Reduce algorithm (TensorGalerkin
  (https://arxiv.org/abs/2602.05052)) that fuses element-wise ops into monolithic GPU kernels.
• JIT-free & debugging-friendly — eager execution, no compilation overhead, no opaque traces.
• Comprehensive element & mesh support — triangle / quad / tet / hex / pyramid / prism, automated mesh generation, Gmsh & VTKHDF5
  I/O.
• Flexible solvers via torch-sla (https://www.torchsla.com/) — linear, nonlinear, and eigenvalue solves across SciPy / native
  PyTorch / CuPy / cuDSS, CPU and GPU.
• Pythonic API — if you can write PyTorch, you can write FEM.

Links

• 📖 Documentation — https://docs.tensor-mesh.com/
• 🖼️ Example gallery — https://docs.tensor-mesh.com/example_gallery/index.html
• 🏠 Homepage — https://www.tensor-mesh.com/

Citation

If you use TensorMesh in your research, please cite the TensorGalerkin paper — see the Citation section
(https://github.com/camlab-ethz/TensorMesh#citation) in the README.