A phase-change switch is a device that flips between an insulating and a conducting state, and it is only as fast as the time it takes to cool back down. When engineers measure that recovery time, they measure one electrical signal from the whole device. That signal is a sum over many nanoscale regions, or domains, each switching back on its own schedule. Genomics has the same problem under a different name. Bulk RNA sequencing averages millions of cells, and a rare subpopulation that drives drug resistance can be invisible in the average. This note proposes taking the computational tools built to undo that biological averaging and applying them to device transients. The goal is to recover how switching latency is distributed across domains in a vanadium dioxide (VO₂) switch, using only its aggregate electrical response.
The question has engineering consequences. A device's repetition rate is set by its slowest domain, not its average one. Suppose the slow thermal tail comes from a few pinned domains. Then the fix is to heat or strain those sites locally. If instead every domain relaxes uniformly slowly, the fix is to reduce thermal mass everywhere. Today's aggregate measurements cannot tell these two cases apart. That is exactly the situation in tamoxifen resistance, where rare stem-like cells cannot be distinguished from a uniform shift of the whole population using bulk assays (Lin et al.).
Core Claim
The decay of conductance after a switching pulse in a small VO₂ device can be modeled as the sum of integer counts of domains that are still metallic. Each domain belongs to one of a few kinetic types, and the measurement is blurred by a known circuit response. If so, an integer-valued deconvolution, anchored by a small reference set of spatially imaged devices, should recover two things that standard relaxation-time fitting cannot: the fraction of slow domains, and the device's tail recovery time. The fitting gap should be largest in the regime that matters, where devices contain only a handful of domains.
Thermal switching latency is the application regime. Heterogeneity masking supplies the method and the evaluation logic. The biology is not claimed to benefit.
The model is
$$G(t) = \Big(h * \sum_{k=1}^{K} g_k\, n_k(\cdot)\Big)(t) + \varepsilon(t),$$
where the terms are defined as follows. $G(t)$ is the measured conductance at time $t$ after the pulse. $n_k(t)$ is the integer number of type-$k$ domains still metallic. $g_k$ is the conductance contributed by one such domain. $h$ is the parasitic circuit kernel, the electrical smearing from the setup, estimated by calibration. $\varepsilon$ is measurement noise. Recovery corresponds to each $n_k$ decaying by unit "death" events, with no new domains switching on.
Prior Work and Its Role
The method comes from Count Bridges. That work models integer data with Poisson birth–death dynamics. It treats unit-level counts as latent variables inside an expectation–maximization loop, which alternates between guessing the hidden counts and refitting the model. It uses projection-guided sampling so that the generated unit profiles sum exactly to an observed bulk measurement. We reuse the birth–death prior and the sum constraint almost directly. The hidden "cells" become domains, and the "bulk count" becomes conductance, divided by $g_k$ and passed through the circuit kernel $h$. The new part is the time dimension and the convolution with $h$, neither of which Count Bridges addresses.
POPPY (Venkat et al.) maps bulk samples onto a single-cell reference atlas. It also shows the limitation we inherit: a cell state missing from the reference cannot be attributed. Our analogue of the atlas is a small set of devices whose domains are imaged optically. Any kinetic type missing from that set will be invisible in our results too. Ha et al. showed that shuffling mutations across inferred subclones degrades therapy-response prediction, which is evidence that the unmixed structure carries real signal. We borrow that shuffle test as a falsification check.
On the device side, three papers motivate the gap:
- Cao et al. showed that strained single-crystal VO₂ nanobeams organize into one-dimensional alternating metal and insulator domains. These domains are countable and optically visible, which is what makes ground truth possible. The same work reported about 10 K of thermal hysteresis and modeled the domains with Cahn–Allen phase-field dynamics, a standard simulation of how coexisting phases grow and shrink. We reuse that simulation to generate synthetic devices with known per-domain latencies.
- Bohaichuk et al. confined heating to a sub-10 nm channel using a carbon nanotube heater. They reached 11 MHz spiking, compared with 0.3 kHz in VO₂-only devices. They also warned that parasitic capacitance in the measurement setup masks the device's intrinsic dynamics, which is why $h$ appears explicitly in our model. Because their device has essentially one channel, it serves as a homogeneous negative control.
- Zalden et al. found that the electronic threshold switch completes in under a picosecond, while the Joule heating that follows drives filamentary crystallization over far longer times. The latency problem therefore lives in the thermal recovery, which is the part we model.
The kHz limits reported by Zheng et al. and the latencies of hundreds of milliseconds reported by Meng et al. are aggregate numbers of exactly the kind we propose to decompose. Haraldsen et al. showed that lattice dynamics lag spin and charge dynamics. That supports treating domain recovery as slow, discrete events rather than a smooth electronic decay.
Research Program
We will build a simulator, an inference method, and a measurement protocol.
Simulator. It extends the Cahn–Allen phase-field model of Cao et al. with a heat-diffusion term and spatially varying pinning: contact regions, strain gradients, and defects. It will generate ensembles of devices with 2 to 50 domains. Each device has known per-domain recovery times and a known slow-domain fraction. Its conductance traces are passed through realistic circuit kernels.
Inference method. It places a birth–death prior over the counts $n_k(t)$, fitted from the imaged reference devices. It samples count trajectories with a projection step so that $h * \sum_k g_k n_k$ matches the measured $G(t)$ within noise. Repeated pulses on the same device provide multiple "bulk samples" that share parameters.
Measurement protocol. Strained VO₂ nanobeams will be pulsed electrically while optical microscopy images their domains. Imaging gives the true count of metallic domains over time. The electrical traces from the same pulses are held out from inference and used only for scoring.
Baselines
The naive reference point is a single-exponential fit to the conductance recovery. Its time constant is what device papers usually report as "the" latency. We will report its error on the tail recovery time, defined as the time until 99% of domains have recovered. Any method that cannot beat this fit on tail error is not worth using.
The strongest existing competitor is distribution-of-relaxation-times inversion. This is the standard impedance-spectroscopy technique: a Tikhonov-regularized inverse Laplace transform that recovers a continuous spectrum of time constants, here applied after deconvolving $h$. A stretched-exponential fit will be reported alongside it, since it is the common one-parameter description of heterogeneous relaxation. No reported result was found on Lacuna for either method applied to VO₂ or phase-change switching transients, so we will establish their numbers ourselves.
Fairness rules apply to all methods:
- every method receives the same traces, the same calibrated $h$, and the same noise estimate;
- regularization strength for the relaxation-time inversion and the prior strength for our method are tuned on the same held-out simulated validation set, with an equal number of hyperparameter trials;
- all methods run on the same CPU budget per trace.
Beating these baselines means lower 1-Wasserstein error on the per-domain latency distribution and lower tail-time error on the imaged real devices, with a confidence interval from bootstrapping devices that excludes zero.
Two ablations isolate the key ingredients. The first replaces the integer birth–death prior with a continuous nonnegative mixture over the same kinetic types, which is the continuous-relaxation analogue of what Count Bridges improved on (Fishman et al.). It tests whether integrality matters. The second removes the imaged reference set and infers kinetic types from the traces alone. It tests whether the "atlas" matters. We expect integrality to help most below roughly ten domains, and to make little difference for large devices.
Why not just…?
Why not just use distribution-of-relaxation-times inversion, which already handles heterogeneous relaxation? This is the real threat. For devices with many domains, a continuous spectrum of time constants approximates a sum of many discrete events well, and we expect the relaxation-time inversion to tie with our method there. Its weakness is structural. It cannot represent "three domains, one of which is slow." Regularization smooths an isolated slow component into a broad shoulder, and a few-domain trace shows discrete steps that a continuous spectrum cannot express. The simulated sweep over domain count is designed to show exactly where the two methods diverge. If no divergence appears, even below ten domains, we will conclude the integer machinery is unnecessary and report that.
Why not just image every device optically and skip inference? Imaging works only for nanobeams that are large enough to resolve optically, on a microscope stage. It does not work for encapsulated photonic or crossbar devices, or for sub-10 nm channels like those of Bohaichuk et al.. Imaging is our ground truth and our reference set, not a deployable diagnostic. The point of the method is to transfer what imaging teaches to devices that can only be measured electrically.
Why not just shrink the active volume, as the carbon-nanotube heater does, and make heterogeneity moot? For single-channel devices this is the right answer, and that device is our negative control. Multi-domain devices, such as strained beams with 30-fold power reduction (Cao et al.) and multilevel memories (Meng et al.), use many domains by design. For those devices, knowing which domains are slow is the actionable question.
Proposed Experiments
Simulation sweep. On synthetic devices, we sweep three factors: domain count, slow-domain fraction (1% to 50%), and circuit kernel width relative to domain recovery time. For each method we report:
- 1-Wasserstein error on the per-domain latency distribution;
- tail-time error;
- area under the curve for detecting whether any slow domain is present.
The last metric mirrors rare-subclone detection.
Real-device test. This uses imaged VO₂ nanobeams with held-out electrical traces. The decisive check is the inferred metallic-domain count over time, compared with the optically counted one.
Two falsification tests.
- Homogeneous control. On carbon-nanotube composite devices, the method must return a single kinetic component. If it reports heterogeneity there, it is fitting noise.
- Shuffle test, after Ha et al.. We permute the domain-type labels in the reference set. If tail-time error does not get worse, the reference is not contributing.
The physical question. Finally, the method is applied to strained beams, as in Cao et al., across temperatures spanning the roughly 10 K hysteresis. The test is whether slow recovery concentrates in a minority of domains, for example those near contacts, or is spread across all of them. A concentrated tail predicts that locally heating the inferred slow sites will shorten recovery more than uniformly heating the whole device with the same energy. A spread tail predicts no such advantage. That intervention is the experiment that would make the inference worth having.
References
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Strain engineering and one-dimensional organization of metal–insulator domains in single-crystal vanadium dioxide beams
- Fast Spiking of a Mott VO2–Carbon Nanotube Composite Device
- Picosecond Electric-Field-Induced Threshold Switching in Phase-Change Materials
- Nonvolatile Electrically Reconfigurable Integrated Photonic Switch Enabled by a Silicon PIN Diode Heater
- Electrical programmable multilevel nonvolatile photonic random-access memory
- Magneto-optical investigation of the field-induced spin-glass-insulator to ferromagnetic-metal transition in the bilayer manganite<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mo>(</mml:mo><mml:msub><mml:mi mathvariant="normal">La</mml:mi><mml:mn>0.4</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">Pr</mml:mi><mml:mn>0.6</mml:mn></mml:msub><mml:mo>)</mml:mo></mml:mrow><mml:mn>1.2</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">Sr</mml:mi><mml:mn>1.8</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">Mn</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">O</mml:mi><mml:mn>7</mml:mn></mml:msub></mml:mrow></mml:math>