Full YAML Reference

This page is generated from nf2.reference and mirrors the public v0.4 YAML schema.

General Run Keys

Key	Type	Default	Description
path	str	./runs/nf2	Directory for checkpoints, logs, and extrapolation_result.nf2.
work_path	str | null	path/work	Optional scratch directory for preprocessed arrays.
logging	dict	{}	Options passed to the Lightning/W&B logger.
meta_path	str	none	Optional previous NF2 state used by series runs.

Data And Geometry

Key	Type	Default	Description
data.geometry	cartesian | spherical	required	Extrapolation geometry.
data.normalization.Mm_per_ds	float	100	Length represented by one model coordinate unit.
data.normalization.Gauss_per_dB	float	1000	Magnetic-field strength represented by one model field unit.
data.boundaries	list[dict]	required	Boundary observations, analytical fields, or maps.
data.boundaries[].Mm_per_pixel	float	dataset default	Spatial sampling for this boundary in Mm per pixel.
data.boundaries[].coordinate_center	list[float] | dict	[0, 0]	Cartesian coordinate assigned to the boundary center, in Mm.
data.validation	list[dict]	geometry default	Validation grids and plotting datasets.
data.sampler	dict	height sampler	Cartesian physics sampling dataset.
data.samplers	list[dict]	random_radial_grouped	Spherical physics sampling datasets.
data.potential_boundary	dict	FFT potential	Cartesian potential boundary data. Use type: none to disable.
data.z_range	list[float]	loader default	Cartesian height range in Mm where supported by the loader.
data.max_radius	float	loader default	Spherical outer radius in solar radii where supported by the loader.
data.iterations	int	10000	Number of random sampler batches per epoch-like pass.
data.num_workers	int	4	Default PyTorch DataLoader workers for training and validation loaders. Series preloading uses this value unless data.data_module_workers is set.
data.train_num_workers	int	data.num_workers	PyTorch DataLoader workers for training loaders.
data.validation_num_workers	int	data.num_workers	PyTorch DataLoader workers for validation loaders.
data.prefetch_factor	int	5	Training DataLoader prefetch factor when workers are enabled.
data.persistent_workers	bool	true	Keep training DataLoader workers alive while a loader is active.
data.preload_data_modules	bool	true	For series runs, preload all step data modules up front instead of loading each step lazily.
data.data_module_workers	int	data.num_workers	Series-only multiprocessing workers used to preload per-step data modules.

Model

Key	Type	Default	Description
model.field	vector_potential | b	vector_potential	Field representation.
model.network.type	siren	siren	Only SIREN networks are supported.
model.network.hidden_dim	int	256 cartesian, 512 spherical	SIREN hidden width.
model.network.layers	int	model default	Number of SIREN layers.
model.network.w0	float	model default	SIREN frequency scale for hidden layers.
model.network.w0_initial	float	model default	SIREN frequency scale for the first layer.

Training

Key	Type	Default	Description
training.epochs	int	15	Number of Lightning epochs.
training.optimizer.start	float	5e-4	Initial learning rate.
training.optimizer.end	float	5e-5	Final learning rate.
training.optimizer.iterations	int	100000	Learning-rate schedule length.
training.reload_dataloaders_every_n_epochs	int	1 for series	Series cadence for advancing to the next dataset.
training.check_val_every_n_epoch	int	1	Lightning validation cadence. Series examples use 10 to validate every 10th dataset.
training.trainer	dict	{}	Additional Lightning Trainer keyword arguments.

Losses And Scaling

Key	Type	Default	Description
losses[].type	str	geometry default	Loss implementation name.
losses[].name	str	type	Stable logging and scaling identifier.
losses[].weight	float | schedule	required when explicit	Loss weight. The legacy lambda key is not accepted.
losses[].datasets	str | list[str]	loss default	Dataset ids used by the loss.
loss_scaling	list[dict]	geometry default	Spatial scaling modules for selected losses.

Callbacks And Transforms

Key	Type	Default	Description
callbacks	list[dict]	geometry default	Validation plots and metrics logged during training.
callbacks[].plot	bool	true	For plotting callbacks, set false to keep scalar metrics but skip image rendering/logging.
transforms	list[dict]	[]	Optional coordinate/field transforms applied to datasets.