dinox is a JAX implementation of Reduced Basis Derivative Informed Neural Operators, built for high performance in single-GPU environments where all training data fits in GPU memory.

The library is designed primarily for PDE learning workflows based on:

• FEniCS 2019.1
• Jacobians computed via hippylib
• Subspace methods provided by bayesflux

dinox provides:

• Reduced basis neural operator architectures
• Derivative-informed training using PDE Jacobians
• GPU-accelerated implementations in JAX and Equinox
• Integration with FEniCS-discretized PDEs

See the hyperelasticity tutorial for a complete walkthrough of the RB-DINO pipeline: problem setup, data generation, training with L2 vs H1 loss, and surrogate evaluation.

Note on tutorial: The tutorial requires access to the files in dinox/examples. Then one can open examples/DINO_Tutorial.ipynb to view the tutorial. We recommend the user perform these steps:

git clone --depth 1 --filter=blob:none --sparse https://github.com/dinoSciML/dinox.git
cd dinox
git sparse-checkout set examples

dinox is central to the repo lazydino_tutorial, which implements a tutorial jupyter notebook for the methods described in the paper: LazyDINO

dinox depends on bayesflux[hippylib], which requires:

• hippylib
• FEniCS 2019.1

FEniCS has system-level dependencies and cannot be installed via pip alone.

You must first create a conda environment with FEniCS 2019.1 before installing dinox.

• NVIDIA driver >= 525 (check with nvidia-smi)
• conda or mamba

Note on CUDA libraries: You do not need to install CUDA Toolkit, cuDNN, or cuSPARSE via conda or your system package manager. The pip wheels for JAX and CuPy bundle their own CUDA 12 runtime libraries. Installing system CUDA alongside pip-bundled CUDA is the most common source of GPU detection failures.

conda create -n fenics-2019.1_env -c conda-forge fenics==2019.1.0 python=3.11
conda activate fenics-2019.1_env

A system-level or conda-set LD_LIBRARY_PATH pointing to a CUDA installation will conflict with the CUDA libraries bundled in the JAX and CuPy pip wheels, causing errors like Unable to load cuSPARSE.

unset LD_LIBRARY_PATH

pip install "jax[cuda12]" cupy-cuda12x nvidia-curand-cu12

# With CuPy GPU support (recommended)
pip install dinox[cupy]

# Without CuPy
pip install dinox

python -c "import jax; print('JAX devices:', jax.devices())"
python -c "import cupy; print('CuPy GPU count:', cupy.cuda.runtime.getDeviceCount())"

You should see your NVIDIA GPU listed. If JAX shows only CpuDevice, check that LD_LIBRARY_PATH is unset (see Step 2).

• Designed for single-GPU workflows where all data fits in GPU memory
• Requires CUDA 12-enabled JAX (pip install "jax[cuda12]") — the pip wheel bundles its own CUDA runtime
• Optional CuPy arrays for GPU operations via dinox[cupy]
• Without GPU, CPU fallback is automatic

conda create -n fenics-2019.1_env -c conda-forge fenics==2019.1.0 python=3.11
conda activate fenics-2019.1_env

pip install "jax[cuda12]" cupy-cuda12x nvidia-curand-cu12
unset LD_LIBRARY_PATH  # or use the permanent conda hook above
pip install -e ".[dev]"

This installs:

• dinox (editable)
• development tools (pytest, black, flake8, isort)
• bayesflux[hippylib]
• hippylib
• all required JAX dependencies

• Python >= 3.10
• FEniCS 2019.1 (via conda)
• JAX >= 0.7.0 (for GPU: pip install "jax[cuda12]")
• NVIDIA driver >= 525 (for GPU)
• Optional: CuPy for GPU array operations (pip install dinox[cupy])

• Homepage: https://github.com/dinoSciML/dinox
• Repository: https://github.com/dinoSciML/dinox