One nonlinear photodetector beats a spectrometer by refusing to be linear

What the work claims

The claim is a paradigm shift stated in one sentence: phenomena such as nonlinearity, hysteresis and contact asymmetry, conventionally treated as defects to be engineered out of a sensor, can become information-bearing degrees of freedom. Conventional spectrometers map wavelength into space with a grating or filter and read the result with an array of linear photodiodes, which is why resolving power trades against footprint. This work replaces the whole dispersive stage with a single planar fluorinated phenethylammonium lead iodide (F-PEAI) photodetector. Wavelength and irradiance modulate not just how much photocurrent flows but the entire shape of the bias-dependent response, through coupled photocarrier generation, trap filling and emptying, interfacial transport barriers, field redistribution and the device's illumination and bias history. A machine-learning inverse model then decodes centre wavelength and irradiance from the full current-voltage trajectory, not from any single measured value.¹

This is a primary experimental result on one material system, not a general proof about arbitrary sensors. Its weight comes from the numbers: the inverse model generalises to excitation wavelengths that were explicitly excluded from training, which is the difference between a lookup table and an encoding.

How it works

The device is deliberately mundane: mechanically exfoliated F-PEAI flakes on prepatterned interdigitated gold electrodes, total width 16.58 mm, channel length 3 um. F-PEAI has a direct bandgap of about 2.61 eV (around 475 nm), strong ultraviolet and blue absorption, a pronounced room-temperature excitonic response with emission centred near 523 nm, and strong sensitivity to its metal-semiconductor interfaces. The authors probe it under monochromatic illumination spanning 370 to 570 nm at 8 nm linewidth, over incident optical powers from 1e-9 to 2.1e-5 W, collecting full forward and reverse bias sweeps of dark current and photocurrent. The spectral responsivity is strongly nonlinear and non-monotonic, peaking at the exciton wavelength of 516 nm.¹

The encoding works because the current-voltage trajectory is a path through device state, not a point measurement. During a sweep the internal state evolves through trap filling and emptying, persistent photoconductivity, carrier relaxation and possible ionic redistribution, so the forward and reverse branches carry different information and their hysteresis is itself a signature of the optical input. Wavelength changes absorption and the microscopic pathways of carrier generation and transport; irradiance changes photocarrier density, trap occupation and the transport regime. Two different optical states that would produce nearly identical photocurrent at one bias can produce clearly distinguishable trajectories across the full sweep.

The decoder is built to respect that structure. Each sweep channel is projected onto a Legendre polynomial basis truncated at degree 8, a representation that preserves global shape, nonlinearity and hysteresis while suppressing point-to-point measurement noise and removing dependence on the particular voltage grid. Forward and reverse sweeps, and dark current versus photocurrent, are encoded separately. A variational encoder maps the coefficients onto a 32-dimensional Gaussian latent space, and a Gaussian decoder emits the physical parameters of the incident spectrum: centre wavelength, amplitude and width per spectral component, with a single component (K = 1) for this monochromatic dataset. The whole model has about 14,800 parameters. Repeated sampling from the latent distribution yields a 100-member prediction ensemble whose spread reports the model's own sensitivity, an internal uncertainty measure distinct from the absolute reconstruction error.¹

The dataset is small and honestly curated: 505 recorded sweeps, 29 discarded by an automated signal-to-noise quality screen, leaving 476, of which 95 form a held-out test set whose wavelength-irradiance pairs were excluded from training. That puts the training set just under 400 sweeps, and the model still reconstructs held-out wavelength with R2 = 0.958, mean absolute error 8.1 nm and RMSE 11.2 nm, while log-normalised irradiance reaches R2 = 0.987 (about 0.96 on the linear scale). The error is not uniform: it concentrates between roughly 425 and 530 nm, exactly where the non-monotonic responsivity makes distinct wavelengths degenerate at a single bias, and the model resolves that region only because it decodes the full trajectory rather than one operating point.¹

Where a skeptic should push

The most load-bearing assumption is stability: the entire method assumes the device's nonlinear, history-dependent response is a fixed encoding that can be calibrated once and inverted forever. The authors themselves hold the measurement temperature and the voltage-sweep protocol fixed, which is an admission that the encoding drifts with conditions. A 2D hybrid perovskite is not a material one would bet a calibration on: trap distributions evolve with light soaking, bias stress and time, and the paper offers no retention or ageing study of the encoding itself. When the physics that carries your information is the same physics that degrades, your calibration and your failure mode share a root cause.

Second, the scale of evidence is one device, one material, monochromatic inputs, and 476 sweeps acquired in a single controlled setup. Generalisation to held-out wavelengths is real, but it is interpolation within one responsivity manifold, not demonstration that the approach survives a different device, a different flake, or broadband spectra; multi-peak reconstruction beyond K = 1 is stated as an architectural capability, with predictive performance explicitly left to future work. Third, the readout cost is understated by the word "single": each estimate requires digitising complete dark and photocurrent sweeps in both directions, tens of bias points per measurement, where a grating spectrometer exposes its array in parallel. The information economics only favour this design when the readout electronics are cheap relative to optics, or when the sensor must be integrated where optics cannot go. All three caveats are fixable; none is addressed in the present data.

When electrode nonlinearity carries information

For microelectrode array work, the paper reads as a provocation about the front end's job description. MEA instrumentation spends most of its design effort doing the opposite of this paper: linearising electrode response, suppressing hysteresis, cancelling stimulation-recovery transients, and treating electrode drift and history effects as enemies of measurement. That is the right instinct when the electrode is meant to be a transparent witness. But this work demonstrates, with measured numbers, that a history-dependent nonlinear interface can carry more independent stimulus dimensions per element than a linear one: their single detector separates two physical quantities (wavelength and irradiance) that a linear photodiode confounds, and it does so after the absolute current magnitude is normalised away, meaning the information lives in the shape of the trajectory, not the amplitude. The specific mechanism is trap occupation dynamics. MEA electrodes have their own slow interfacial state: charge redistribution in the electric double layer, surface chemistry drift, and conditioning history after stimulation. Every one of those is currently a calibration liability. This paper supplies the template for asking when they could instead be a sensing channel: full per-electrode impedance or stimulation-recovery trajectories, digitised and inverted with a learned model, could in principle encode variables no linear channel reports, from local chemical state to interface age, at the cost of the multi-point measurement protocol the authors actually use.

The opportunity for the acquisition chain is bandwidth accounting. Arrays face an egress wall: thousands of channels of wideband data cannot all leave the chip. This paper shows a concrete information-theoretic trade at the device level: a deliberately nonlinear sensor plus an inverse model exchanges raw measurement points for stimulus separability. Transposed to the array, the analog is encoding stimulus dimensions into structured electrode excitations rather than streaming raw voltage, letting device physics do multiplexing that silicon would otherwise pay for in ADCs and pins.

The threat deserves equal weight. If part of an instrument's measurement claim lives in a learned inverse of nonlinear electrode physics, then electrode ageing is no longer a slow nuisance but an uncontrolled re-parameterisation of the model's input mapping, and the 425 to 530 nm ambiguity region in this paper is the cautionary shape: wherever the encoding degenerates, the inverse model still emits a confident number. The authors' own safeguard, an ensemble spread as an internal uncertainty flag, only works if the drift stays inside the calibrated manifold. For arrays, where electrodes age across weeks of culture and stimulation, that assumption is the whole ballgame.

The bottom line

Established: a single nonlinear F-PEAI photodetector, read by full bidirectional current-voltage sweeps and decoded by a roughly 15,000-parameter variational model, reconstructs held-out monochromatic wavelength to a mean absolute error of 8.1 nm and log-normalised irradiance to R2 = 0.987, with error localised to the responsivity-degenerate band by a mechanism the authors identify and explain. Not established: stability of that encoding over time, transfer across devices or materials, or spectra beyond single peaks; the honest scope is one well-characterised sensor. What would confirm the paradigm is the same experiment run against a deliberately aged or thermally cycled device, showing the learned encoding survives re-calibration; what would break it is evidence that trap-state evolution outpaces any practical recalibration cadence. For MEA instrumentation the result is a reframing rather than a product: it is measured proof that the nonlinearities array engineers suppress can be computational assets, and a concrete warning that the asset and the failure mode are the same physics.