A quantum dot is a nanoscale semiconductor island that behaves like an artificial atom. When one is placed in an optical cavity, a tiny mirror structure that traps light, it can still pour energy into the cavity even when their colors do not match. Vibrations of the crystal lattice, called phonons, absorb or supply the energy difference. This "incoherent" feeding is a nuisance and a resource at once. It sets how well a dot and cavity can talk, yet every dot comes out of growth with a different color, so each chip must be characterized one device at a time. That characterization is slow. It needs sweeps over temperature, laser power and detuning, followed by careful model fitting.
We propose to treat this as a diagnostics problem in the ML sense. A physics simulator that can be differentiated, meaning gradients flow through it the way they do through a neural network, is trained offline. It then reads a single short measurement of cavity emission and returns, in well under a second, calibrated estimates of each device's detuning, phonon coupling strength, pure dephasing and local temperature. The same differentiable model chooses the two or three laser settings that carry the most information, so that slow sweeps are no longer needed. The lesson borrowed from rapid clinical diagnostics is a specific failure mode. Learned spectral classifiers break on inputs outside their training range, and we test whether building the physics into the inference fixes that.
Core Claim
The ratio of uphill to downhill phonon-assisted feeding is fixed by detailed balance. Detailed balance is the thermodynamic rule that absorbing a phonon is rarer than emitting one by a Boltzmann factor:
$$\frac{\Gamma_{\uparrow}(\Delta)}{\Gamma_{\downarrow}(\Delta)} = e^{-\hbar|\Delta|/k_B T},$$
where $\Gamma_{\uparrow}$ and $\Gamma_{\downarrow}$ are the cavity feeding rates when the cavity sits above or below the dot, $\Delta$ is the dot–cavity detuning, and $T$ is the lattice temperature. Because of this built-in thermometer, a handful of drive settings can make the phonon spectral density, the detuning and the temperature separately identifiable. Without the incoherent channel these quantities are mixed together in a single Lorentzian peak.
We claim three things. First, an amortized posterior estimator, a network trained once on simulations that outputs a probability distribution over device parameters for any new spectrum, can match per-device least-squares fitting in accuracy at roughly a thousandfold lower cost per device. Second, a measurement plan chosen by gradient-based experimental design beats a uniform sweep that uses the same number of acquisitions. Third, because the physics is built in, the estimator degrades gracefully outside the training range of detuning and temperature, where a black-box spectral network fails.
Grounding in Prior Work
The physical model comes from experiments on single InAs dots in photonic crystal cavities. Those experiments established downhill and temperature-dependent uphill feeding under resonant dot excitation (Phonon mediated off-resonant quantum dot–cavity coupling). They showed that the dot and cavity linewidths broaden with drive power through this channel (Linewidth broadening of a quantum dot coupled to an off-resonant cavity). They also showed that off-resonant cavity emission still reveals the coherent Mollow triplet, the three-peaked spectrum of a strongly driven emitter (Probing of single quantum dot dressed states via an off-resonant cavity). Together, these papers tell us that one cavity spectrum mixes coherent and incoherent information, which is why it is rich enough to diagnose from.
Two further results set the scale of the problem. Spectral mismatches up to 1.2 THz were bridged between two dots through an off-resonant cavity (Phonon-mediated coupling between quantum dots through an off-resonant microcavity). Dot–supermode anticrossings were observed in coupled cavities (Cavity quantum electrodynamics with a single quantum dot coupled to a photonic molecule). Both define the parameter range and the harder two-mode test case. None of these works addresses throughput; each characterizes a handful of devices by hand.
The computational method comes from Driving enhanced exciton transfer by automatic differentiation. That work differentiates a Lindblad master equation, the standard equation for a quantum system leaking energy to its environment, in PyTorch, and uses Adam to optimize driving fields for transport. It reports more than an order-of-magnitude gain in transfer efficiency over the undriven case, and finds that a few driving terms usually suffice. We reuse the machinery but change the objective. Instead of maximizing the energy delivered to a target site, we maximize the information a measurement carries about unknown device parameters. That paper's finding that a few drive terms suffice motivates our bet that two or three well-chosen acquisitions suffice.
The diagnostics literature constrains the design rather than supplying hardware. Two SERS platforms, which classify Raman spectra from silver nanorod substrates with CNNs, report high in-distribution accuracy and degraded performance on concentrations outside training: CoVari (Rapid Detection of SARS-CoV-2 Variants Using an ACE2-Based SERS Sensor) and MultiplexCR (Multiplex Detection and Quantification of Virus Co-Infections). MultiplexCR drops from 98.6% test accuracy to 95.8% on blind specimens. Their weakness is the one we target.
We deliberately do not propose a quantum-dot clinical biosensor. The phonon-feeding and Mollow experiments run near 4 K, where pure dephasing is suppressed. At body temperature in saliva the coherent features this physics relies on are washed out. A passive room-temperature cavity sensor already reaches 0.05 ng/mL by tracking refractive-index shifts (Optical Biosensor Based on Porous Silicon and Tamm Plasmon Polariton). An emitter-fed cavity would trade that sharp reflection dip for a broad phonon sideband, and we see no case that it would win.
Research Program
We will build a differentiable simulator in PyTorch for a driven two-level dot coupled to one cavity mode, and later two modes. The simulator includes phonon-assisted feeding terms whose rates follow a parametrized phonon spectral density and Bose occupation. It also includes pure dephasing and cavity loss, and computes steady-state cavity emission spectra and their power dependence. A second version uses a polaron-frame treatment, which handles strong electron–phonon coupling more accurately. This version serves as a deliberately mismatched "true" generator, so that we can test robustness to the Markovian simplifications the field acknowledges.
On top of the simulator we will do two things. First, we train a normalizing-flow posterior estimator, a network that outputs a full distribution over the physical parameters given a few spectra. Training draws device parameters from priors set by the published regimes and adds realistic detector noise. Second, we solve an experimental-design problem. We choose $K$ drive settings $u_1,\dots,u_K$, each a combination of laser power, laser detuning and stage temperature, to maximize the expected log-determinant of the Fisher information. The Fisher information measures how sharply the data pin down the parameters, and its gradients are obtained by differentiating through the simulator. The deliverable is a screening protocol that labels each device as usable or not, for example by whether its incoherent-feeding fraction and dephasing fall within thresholds, together with calibrated uncertainty.
Baselines
The naive reference fits each spectrum with independent Lorentzians and reads detuning from peak positions and coupling from intensity ratios. This is close to how quick device sorting is done by hand. It shows what is lost when phonons are ignored.
The strongest competitor is per-device nonlinear least-squares or Bayesian fitting of the same simulator, run with multiple random restarts, on the same measured spectra. It uses the identical physical model, so it isolates the value of amortization and of measurement design rather than of the physics. We give it the same acquisitions plus, separately, a full dense sweep, and report the accuracy versus wall-clock frontier.
We also train a black-box CNN regressor in the style of MultiplexCR and CoVari on the same simulated spectra, with matched parameter count and tuning budget. This tests whether physics-built-in inference extrapolates better.
The key ablation removes the phonon-feeding terms, leaving a cavity-QED model with only pure dephasing. This shows whether the detailed-balance asymmetry is what makes temperature and coupling identifiable.
No reported result on Lacuna exists for parameter-inference accuracy or throughput on quantum dot–cavity devices. The only reported figure is the 1.2 THz bridged mismatch, which sets our detuning prior range. Beating the strongest baseline means matching its parameter error within its own uncertainty and its usable/unusable sorting accuracy, at under one second per device. It also means reaching that accuracy with a designed three-setting measurement where the fitting baseline needs a uniform sweep at least several times longer.
Why not just…?
Why not just fit every device with the physical model? Fitting is the accuracy gold standard, and it is a real threat to our claim. If multi-start fitting from the three designed acquisitions already runs in about a second with equal accuracy, the amortization contribution collapses. What remains is the experimental design. The head-to-head frontier experiment settles this directly. We expect fitting to suffer from multimodal likelihoods near zero detuning and at high dephasing, and we will report its restart-failure rate.
Why not just train a CNN on spectra, as the SERS diagnostics do? It is simpler and fast at inference. But both CoVari and MultiplexCR document degradation outside the training distribution, and here new wafers routinely bring new detunings and temperatures. Our out-of-range split tests whether the physics-built-in estimator holds up where the CNN does not. If the CNN extrapolates equally well, we will say so.
Why not just run the full temperature and power sweep? It is reliable but costs much more cryostat time per device, and throughput is the bottleneck that random dot-to-dot detuning creates for scaling multi-emitter chips. The designed-versus-uniform comparison at equal acquisition count measures how much of the sweep is redundant.
Why not build the room-temperature clinical sensor directly? The physics argues against it, for the reasons above. The Tamm-plasmon sensor already does the passive version well. We redirect the diagnostics motivation to where incoherent coupling physics genuinely gives an advantage.
Proposed Experiments
The first experiment is a synthetic benchmark. We draw ten thousand simulated devices spanning detunings to 1.2 THz, temperatures from 4 to 40 K and dephasing over two decades. We report parameter error, posterior coverage at the 50% and 90% levels (a check that the uncertainty bars are honest), sorting accuracy, and wall-clock time per device for every baseline.
The second is a held-out extrapolation split, in which training covers detunings below 0.6 THz and temperatures below 20 K, and testing covers the rest. This is the decisive check against the black-box CNN.
The third is a misspecification test, generating data with the polaron model and inferring with the Lindblad model. This checks whether calibration survives non-Markovian effects or whether the estimator becomes confidently wrong.
The fourth repeats the pipeline on the two-mode photonic-molecule model to test scaling.
If a collaborating lab can supply spectra from devices like those in the cited experiments, the final check compares our three-acquisition estimates against their full-sweep fits on the same devices. Without such data, the claims stay at the simulation level, and we will state them that way.
References
- Phonon mediated off-resonant quantum dot–cavity coupling under resonant excitation of the quantum dot
- Linewidth broadening of a quantum dot coupled to an off-resonant cavity
- Probing of single quantum dot dressed states via an off-resonant cavity
- Phonon-mediated coupling between quantum dots through an off-resonant microcavity
- Cavity quantum electrodynamics with a single quantum dot coupled to a photonic molecule
- Driving enhanced exciton transfer by automatic differentiation
- Rapid Detection of SARS-CoV-2 Variants Using an Angiotensin-Converting Enzyme 2-Based Surface-Enhanced Raman Spectroscopy Sensor Enhanced by CoVari Deep Learning Algorithms
- Multiplex Detection and Quantification of Virus Co-Infections Using Label-free Surface-Enhanced Raman Spectroscopy and Deep Learning Algorithms
- Optical Biosensor Based on Porous Silicon and Tamm Plasmon Polariton for Detection of CagA Antigen of Helicobacter pylori