Photodynamic model of a quantum defect

Testing 101 models of how a spin defect in hBN glows gave a better fit than both published ones.

beats published

Simplified level diagram of the eleven-level model the campaign identified (additions in teal), and the search score across the 101 model structures tested, dropping below the published nine-level model toward the counting-noise floor. — Measured data · diagram simplified from the preprint's Fig. 3b; plot from the campaign's records (Fig. 3c)

101model structures fit to measured data

0.83 vs 2.13fit error (SSR, smoothed traces) vs published nine-level

6best-so-far fits set aside after checks

The problem

The boron vacancy (V_B⁻) in hexagonal boron nitride is a candidate 2D quantum sensor whose glow depends on its spin. Fitted rates disagree across studies; extra levels add free parameters, so a better fit alone isn't enough.

What we did

Starting from our published nine-level model, with scoring fixed in advance (one joint fit to measured photoluminescence at five laser powers), the collaborative agent fit 101 level schemes and tested promising ones on withheld powers. Researchers steered.

Result

An eleven-level model: one-photon entry into a ladder of three dark states, paths back to the ground state, and a light-driven shelf state. Fit error (SSR, smoothed traces) 0.83, vs 2.13 and 3.44 for the published nine- and seven-level models.

Keep in mind

Fit isn't proof: extra levels add free parameters, and the withheld-power comparison with the published seven-level model is being re-checked; two added states have no ab initio counterpart yet.

In the preprint

Sec. IV.B, “Improved photodynamic model for V_B⁻ spin defects in hBN,” and Fig. 3 of arXiv:2610.17193 describe the campaign and its checks in more detail.

Fig. 3 of the preprint, full size — Fig. 3 of the preprint: (a) the published nine-level model; (b) the model identified in the campaign, with additions and changes in teal; (c) the search score (fractional SSR, a different scale from the SSRs quoted above) of each of the 101 model structures, in the order tested, with the published seven- and nine-level models and the counting-noise floor. Click to enlarge.

Data: Measured data
Models (as stated in the preprint): GPT 6 Astra
PhysKit version (as stated in the preprint): 0.12.1, an earlier version
Where: Sec. IV.B, Fig. 3

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.