nf2 0.4.0

Neural non-linear force-free magnetic field extrapolation framework

Neural Network Force-Free Magnetic Field Extrapolation - NF2

---

NF2 is a Python framework for neural non-linear force-free magnetic-field extrapolations. It supports Cartesian and spherical geometries, single-boundary and multi-height observations, extrapolation series, standard NLFF quality metrics, and exports for scientific analysis and visualization.

The framework is designed for solar-physics analyses with HMI/SHARP, full-disk, synoptic, and benchmark data, while keeping the training, export, and evaluation interfaces consistent across geometries.

Installation

NF2 supports Python 3.11 and 3.12.

Install from PyPI:

pip install nf2

Install with conda:

conda install nf2

For GPU-enabled PyTorch installs, select your CUDA version at pytorch.org/get-started/locally. For CUDA 12.6, run:

pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu126

For development or source installs:

git clone https://github.com/RobertJaro/NF2.git
cd NF2
python -m pip install -r requirements.txt

The recommended conda environment for local development is:

conda env create -f environment.yml
conda activate nf2

See the online installation guide for pip, conda, local, and development installation options.

Quick Start

Run an extrapolation from a YAML configuration:

nf2-extrapolate \
  --config examples/configs/cartesian/sharp_cea.yaml \
  --run_path "./runs/ar377" \
  --work_path "/scratch/ar377" \
  --Br "/data/Br.fits" \
  --Bt "/data/Bt.fits" \
  --Bp "/data/Bp.fits"

Download HMI SHARP data:

nf2-download \
  --source hmi_sharp \
  --download_dir "./data/sharp" \
  --email "[e-mail]" \
  --sharp_num 377 \
  --t_start "2011-02-15T00:00:00"

Export a trained extrapolation:

nf2-export \
  "path/to/extrapolation_result.nf2" \
  --format vtk \
  --out "path/to/field.vtk" \
  --Mm_per_pixel 0.72 \
  --metrics j alpha free_energy_fft

Print standard NLFF quality metrics:

nf2-metrics \
  "path/to/extrapolation_result.nf2" \
  --Mm_per_pixel 0.72 \
  --height_range 0 80

Run a time series or parameter series:

nf2-extrapolate-series --config "path/to/series.yaml"

Python API

import nf2

out = nf2.load("path/to/extrapolation_result.nf2")
cube = out.load_cube(Mm_per_pixel=0.72, metrics=["j", "alpha"])

j is the current-density magnitude |J|. Request j_vec when you need the full current-density vector, for example for custom force-free alignment metrics.

Direct helpers are also available:

from nf2 import CartesianOutput, SphericalOutput, run, run_series

Examples

Example configurations live in examples/configs; see the examples overview for guidance by run type:

• cartesian/sharp_cea.yaml
• cartesian/minimal_fits.yaml
• cartesian/auto_disambiguation.yaml
• cartesian/multi_height.yaml
• cartesian/sharp_cea_series.yaml
• cartesian/multi_height_series.yaml
• spherical/hmi_full_disk.yaml
• benchmark/analytical_case1.yaml
• benchmark/analytical_case2.yaml

Notebooks are available in examples/notebooks, including a Colab SHARP CEA tutorial, local SHARP CEA, Cartesian series, spherical HMI, and analytical benchmark runs. Command-line examples for downloads, extrapolations, exports, metrics, and series runs are split by topic under examples/scripts.

Documentation

The full documentation is online at nf2.readthedocs.io.

The documentation includes:

• Installation instructions for pip, conda, local checkouts, and development setups
• Quickstart and usage overview
• YAML configuration reference
• Cartesian, spherical, analytical, and series extrapolation runs
• Training guidance for losses, schedules, height scaling, validation resolution, and memory use
• Evaluation guidance for exports and quality metrics
• Dataset, sampler, and normalization options
• Download, extrapolation, export, and metric commands
• Python API reference
• Example notebook descriptions
• FAQ
• Publication list

Visualization

NF2 results can be exported to VTK and visualized with ParaView:

Publications

Core NF2 method and tool-development papers:

• Jarolim, Thalmann, Veronig, and Podladchikova 2023, Nature Astronomy
  "Probing the solar coronal magnetic field with physics-informed neural networks."
  DOI: 10.1038/s41550-023-02030-9
• Jarolim, Tremblay, Rempel, Molnar, Veronig, Thalmann, and Podladchikova 2024, The Astrophysical Journal Letters
  "Advancing solar magnetic field extrapolations through multiheight magnetic field measurements."
  DOI: 10.3847/2041-8213/ad2450
• Jarolim, Veronig, Purkhart, Zhang, and Rempel 2024, The Astrophysical Journal Letters
  "Magnetic field evolution of the solar active region 13664."
  DOI: 10.3847/2041-8213/ad8914
• da Silva Santos, Dunnington, Jarolim, Danilovic, and Criscuoli 2025, The Astrophysical Journal
  "Magnetic Reconnection in a Compact Magnetic Dome: Chromospheric Emissions and High-velocity Plasma Flows."
  DOI: 10.3847/1538-4357/adcf23

See docs/publications.md for selected applications.

License

NF2 is released under the GPL-3.0 license.