FAQ

How can I validate the quality of the extrapolations?

Use validation callbacks during training and run nf2-metrics after training. The most useful checks are divergence metrics, current-field alignment metrics such as theta_J and sigma_J, force-free residuals, magnetic-energy values, and visual inspection of boundary/slice plots.

nf2-metrics ./runs/case/extrapolation_result.nf2 --Mm_per_pixel 0.72 --height_range 0 80

Also compare exported fields and derived quantities such as j, alpha, and free energy over the region where the boundary data are reliable.

How do I handle Out Of Memory errors?

Reduce the sampler and validation load before changing the model. Lower data.sampler.batch_size, boundary batch_size, validation batch_size, and validation resolution. For Cartesian runs, increasing data.potential_boundary.strides reduces potential-boundary memory. For post-training evaluation, pass a smaller --batch_size to nf2-metrics.

See Training for a step-by-step order.

When is the extrapolation run converged?

Treat convergence as a combination of stable losses, stable quality metrics, and physically plausible validation plots. Boundary losses should stop improving rapidly, force-free and divergence metrics should flatten, and validation slices should no longer show large structural changes between validation intervals. For series runs, check that each step behaves consistently rather than judging only the final step.

If metrics are still improving at the end of training, increase training.epochs or data.iterations. If only one loss is dominating, tune its weight or schedule.

How do I avoid using W&B online?

Set W&B to offline or disabled mode in the environment before launching NF2. Offline mode keeps local logs without syncing:

WANDB_MODE=offline nf2-extrapolate --config nf2/cartesian/sharp_cea.yaml --run_path ./runs/case

Disabled mode turns W&B logging off more aggressively:

WANDB_MODE=disabled nf2-extrapolate --config nf2/cartesian/sharp_cea.yaml --run_path ./runs/case

You can still keep a minimal logging block in the YAML for run names and local organization.