Current-density reconstruction

Mapping currents from noisy magnetic images with a third less test error than our published network.

−33% error · synthetic

A typical synthetic test image reconstructed by the published network (broken segments) and by the campaign's network (follows the wire pattern), and the mean test error over the 34 experiments plotted in the preprint's Fig. 2a, ending 33% below the published network. — Synthetic test data · a typical test image · maps recomputed with the campaign's saved networks; plot from the campaign's records (preprint Fig. 2)

≈33%lower test error than published network

3.9 M vs 15.5 Mtrainable parameters (¼ as many)

12 GBsingle consumer GPU for the campaign

The problem

Turning a magnetic image back into the currents that made it is an inverse problem: direct Biot–Savart inversion amplifies noise without bound. The bar to beat: MAGIC-UNet, our network for quantum diamond microscope images.

What we did

Researchers set the goal (lower average error on synthetic test images at four noise levels) and steered. The collaborative agent reproduced our published results, then searched networks and training, checking gains against seed-to-seed spread.

Result

OPTI-Net, a U-Net built to respect the pixel grid's 8 rotations and reflections and trained across noise levels: about 33% lower error on synthetic test data, ¼ the parameters. Physics-based add-ons didn't help.

Keep in mind

Scored on synthetic test data, averaged over four noise levels. Real microscope images have no ground truth, so no gain can be scored on them.

In the preprint

Sec. IV.A, “Current density reconstruction,” and Fig. 2 of arXiv:2610.17193 describe the campaign and its checks in more detail.

Fig. 2 of the preprint, full size — Fig. 2 of the preprint: (a) the campaign's test error (RMSE) at four noise levels and their mean for each experiment, new bests starred, with the published network (white diamond) for comparison; (b) two example synthetic test images chosen in the preprint, at the highest noise level (SNR = 1): the noisy input B_z, the ground truth, and the reconstructions by the published network and by the campaign's network. Click to enlarge.

Data: Synthetic test data
Models (as stated in the preprint): Claude Opus 4.8
PhysKit version (as stated in the preprint): 0.6.0, an earlier version
Where: Sec. IV.A, Fig. 2

We list the versions the preprint reports so readers can trace each result to its tools, not as a recommendation: for new work, use the latest release.