Repetitive Non-Destructive Readout (RNDR)

• Repetitive Non-Destructive Readout (RNDR) is a technique that measures the same stored signal repeatedly, reducing noise approximately as 1/√N while preserving the state.
• Various implementations, such as SiSeRO, Skipper-in-CMOS, and RNDR-DEPFET, employ unique architectures to maintain charge integrity and enable multiple read cycles.
• RNDR finds applications in low-noise X-ray spectroscopy, quantum computing, and advanced memory devices, addressing challenges in state preservation and efficient information extraction.

Repetitive Non-Destructive Readout (RNDR) is a measurement strategy in which the same stored signal is interrogated multiple times without being destroyed. In detector implementations, the stored object is typically a pixel charge packet; in other realizations it is a programmed ferroelectric polarization state, a qubit state, or a trapped-particle population. The defining operational feature is repeated access to the same physical state, and the defining statistical advantage in charge-sensing systems is that, if each read has independent noise $\sigma$, averaging $N$ repeated measurements reduces the effective noise to $\sigma_N=\sigma/\sqrt{N}$ (Chattopadhyay et al., 2023). In quantum-information settings, closely related quantum non-demolition and low-loss readout protocols use repetition to extinguish measurement errors while preserving the state for reuse (Yoneda et al., 2019, Nikolov et al., 2023). This suggests that RNDR is best understood as a protocol class spanning detector physics, memory devices, and qubit measurement, rather than as a single device architecture.

1. Core measurement principle

The canonical RNDR formulation in semiconductor detectors is explicit: the same charge packet is measured multiple times without being consumed by the measurement, and independent noise averages down approximately as $1/\sqrt{N}$ (Chattopadhyay et al., 2023). In the SiSeRO literature, this is written as

$$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$

while related Skipper-in-CMOS, RNDR-DEPFET, and MAS-CCD work uses the same scaling law for repeated sampling of an unchanged signal charge (Lapi et al., 2024, Wernicke et al., 17 Nov 2025, Botti et al., 14 Feb 2025).

The principle generalizes when repeated measurements are distributed across multiple sensors or amplifiers. In the Multiple-Amplifier Sensing CCD (MAS-CCD), the final pixel value is defined as

$$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$

with effective noise

$\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$

so the square-root gain is extended over both sample count and amplifier count (Botti et al., 14 Feb 2025). In qubit platforms, the same logic appears in a different form: repeated QND-compatible measurements suppress assignment error because the state remains correlated across cycles, provided back-action is sufficiently small (Yoneda et al., 2019).

A common misconception is that “non-destructive” always means “perfectly disturbance-free.” The literature is more precise. In SiSeRO and Skipper-like charge detectors, the charge packet is intended to remain intact during sensing; in neutral-atom and molecular systems, the atom or molecule usually survives the measurement with high probability but not with unit probability; and in superconducting NDRO memory, the stored state is restored by local feedback after a destructive read event rather than sensed in a strictly passive manner (Radnaev et al., 2024, Salas-Estrada et al., 1 Jun 2026, Ucpinar et al., 2023).

2. SiSeRO and X-ray CCD implementations

SiSeRO (Single electron Sensitive Readout) provides one of the most explicit RNDR realizations in X-ray imaging. The architecture uses a p-MOSFET transistor with a depleted internal gate beneath the transistor channel. Charge transferred into the internal gate modulates the source-drain current, so the signal is sensed as a drain-current change rather than by dumping charge onto a destructive floating diffusion. The 2021 and 2022 SiSeRO papers emphasize that the charge packet in the internal gate is unaffected in the readout process and can be moved around like any charge packet in a CCD, making RNDR structurally compatible with the device from the outset (Chattopadhyay et al., 2021, Chattopadhyay et al., 2022).

In the first SiSeRO RNDR demonstrations, the charge packet was shuttled non-destructively between the internal gate and the adjacent summing well (SW) through output-gate (OG) clocking. In the 2023 demonstration, normal non-RNDR operation held OG at about $0.5$ V, while RNDR clocked it between $\mathrm{OGLow}\approx 0.5$ V and $\mathrm{OGHigh}\approx 4$ V; SW and OG were then sequenced to create alternating baseline and signal regions in the video waveform, with correlated double sampling applied cycle by cycle (Chattopadhyay et al., 2023). An earlier Stanford characterization paper reported preliminary 9-cycle RNDR with an exponent $N$0, about $N$1, and about $N$2 eV FWHM at $N$3 keV (Chattopadhyay et al., 2022). The formal first demonstration on a prototype MIT Lincoln Laboratory CCID-93 buried-channel device reported a single-cycle read noise of $N$4, improving to $N$5 after nine cycles, with a fitted exponent $N$6, and a $N$7 keV FWHM improving from $N$8 eV in the first cycle to $N$9 eV after the ninth cycle (Chattopadhyay et al., 2023).

The same device class was subsequently driven into the sub-electron regime. Using an enhanced setup with operation at $\sigma_N=\sigma/\sqrt{N}$0C, improved circuitry, and advanced digital filtering, a 2024 SiSeRO study implemented $\sigma_N=\sigma/\sqrt{N}$1 RNDR cycles of about $\sigma_N=\sigma/\sqrt{N}$2s CDS each and reduced the ENC from $\sigma_N=\sigma/\sqrt{N}$3 in the first cycle to $\sigma_N=\sigma/\sqrt{N}$4 in the $\sigma_N=\sigma/\sqrt{N}$5th cycle, while the centroid of the $\sigma_N=\sigma/\sqrt{N}$6 keV line remained constant over all cycles, implying no charge loss during $\sigma_N=\sigma/\sqrt{N}$7 repetitive transfers (Chattopadhyay et al., 2024). A 2025 electronics paper then reported $\sigma_N=\sigma/\sqrt{N}$8 ENC after $\sigma_N=\sigma/\sqrt{N}$9 RNDR cycles and framed RNDR as a core capability of SiSeRO-based CCDs and prospective active pixel sensors, alongside an 8-channel ASIC denoted the Multi-Channel Readout Chip (MCRC) with an experimental drain current readout mode for SiSeRO devices (Chattopadhyay et al., 19 Aug 2025).

The SiSeRO sequence also makes the main RNDR tradeoff unusually clear. Nine cycles at about $1/\sqrt{N}$0 kpixel/s per cycle corresponded to an effective overall rate of about $1/\sqrt{N}$1 kpixel/s in the 2023 demonstration (Chattopadhyay et al., 2023). Moreover, the read noise continued to fall approximately as $1/\sqrt{N}$2, but the X-ray spectral FWHM stopped improving beyond about the fourth cycle because thermal dark current began to dominate the total noise budget in the available $1/\sqrt{N}$3C “Tiny Box” setup (Chattopadhyay et al., 2023). The later $1/\sqrt{N}$4C measurements were therefore not merely incremental; they addressed the specific dark-current bottleneck that had prevented additional RNDR cycles from translating directly into better spectroscopy (Chattopadhyay et al., 2024).

For X-ray astronomy, the significance claimed for RNDR-enabled SiSeRO is twofold: very low noise with preserved full signal range, and sub-electron sensitivity that can support in-situ absolute calibration of gain and low-energy response (Chattopadhyay et al., 2023). The forward program described in the SiSeRO literature includes RNDR-optimized structures with two adjacent SiSeRO transistors, CCDs with 16 SiSeRO amplifiers in series, and an $1/\sqrt{N}$5 SiSeRO active pixel sensor with two SiSeRO amplifiers per pixel (Chattopadhyay et al., 19 Aug 2025, Chattopadhyay et al., 2024).

3. Semiconductor RNDR beyond SiSeRO

RNDR has also been realized in several Skipper-derived and DEPFET-derived semiconductor architectures. In Skipper-in-CMOS, a CMOS pixel embeds a Skipper-CCD-like output stage with a pinned photodiode, summing gate, sense node, dump gate, reset transistor, and source follower, so that charge can be moved between SG and SN, measured, moved back, and finally dumped. The measured noise followed the expected $1/\sqrt{N}$6 law up to $1/\sqrt{N}$7, reaching $1/\sqrt{N}$8 in the noise scan and $1/\sqrt{N}$9 from the fit to the $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$0 peak, with first results obtained from a $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$1 pixel cell fabricated in Tower Semiconductor’s commercial 180 nm CMOS Image Sensor process (Lapi et al., 2024).

MAS-CCD extends Skipper-style RNDR by placing multiple floating-gate amplifiers in series, so that the same charge packet can be measured repeatedly by different amplifiers as it propagates through the serial register. The architecture was demonstrated in an 8-amplifier and a 16-amplifier sensor. In the 16-amplifier continuous-readout mode, combining all 16 channels yielded a mean pixel noise around $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$2 ADU, corresponding to about $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$3 electron RMS, at $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$4 pixels/s. In the 8-amplifier region-of-interest mode, using $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$5 non-destructive measurements per amplifier in a $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$6 pixel ROI with 7 working amplifiers gave $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$7 and made a weak projected object with an average signal around $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$8pixel clearly visible (Botti et al., 14 Feb 2025).

RNDR-DEPFET uses a different local mechanism. Each pixel contains two DEPFET sub-pixels connected by a transfer gate, so electrons stored in one internal gate can be read out, shifted to the other DEPFET sub-pixel, read again, and cycled repeatedly. In the DANAE prototype, this was implemented in a $$\sigma_N = \frac{\sigma}{\sqrt{N_{\mathrm{cycle}}}},$$9 detector with $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$0 pixel size, $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$1 thickness, $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$2 mg sensor mass, and $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$3 repetitions per pixel, enabling electron-number resolution and a fitted charge carrier generation rate of

$$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$4

The paper emphasizes sensitivity to rare events with $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$5 or more electrons, with events with $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$6 electrons less than $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$7 and events with $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$8 or $$\text{pixel value} = \frac{1}{N_a N_s}\sum_{j=1}^{N_a}\sum_{i=1}^{N_s} s_{j,i},$$9 electrons less than $\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$0 (Wernicke et al., 17 Nov 2025).

Platform	Non-destructive mechanism	Reported result
Skipper-in-CMOS	Repeated SG/SN sampling of the same packet	$\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$1 rms at $\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$2
MAS-CCD	Multiple floating-gate amplifiers in series	$\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$3 in ROI mode
RNDR-DEPFET	Transfer between two readout nodes in each pixel	$\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$4 repetitions with electron-number resolution
SiSeRO	Shuttling between internal gate and SW/OG	$\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$5 after $\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$6 cycles

Taken together, these results show a consistent architectural theme: RNDR is enabled either by a floating-gate or internal-gate sensor that does not consume the stored charge during readout, or by an in-pixel transfer topology that keeps the same electrons available to a second sensing node. This suggests that the main design variable is not the averaging law itself, which is generic, but the device-level mechanism that preserves state fidelity while keeping throughput acceptable.

4. Non-destructive readout in memory devices

In ferroelectric HfO$\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$7/ZrO$\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$8 capacitive memories, RNDR is defined as a read scheme that can be applied many times without changing the programmed polarization state while still allowing the state to be distinguished electrically. The device is a BEOL-compatible FeCap with a $\sigma = \frac{\sigma_0}{\sqrt{N_s}\sqrt{N_a}},$9 nm $0.5$0 ferroelectric nanolaminate grown by ALD on $0.5$1, with a W bottom electrode and a TiN top electrode, integrated both on thermal $0.5$2 and in XFAB 180 nm CMOS. Partial switching creates MemCapacitance states that can be read non-destructively at $0.5$3 V for one month, with projected $0.5$4 MC ratio after $0.5$5 years and endurance up to $0.5$6 cycles at $0.5$7 V (Baigol et al., 2 Jun 2026).

That conventional non-destructive C–V readout does not remain valid at high speed. The MC window shrinks and collapses above about $0.5$8 MHz, summarized in the abstract as a limitation above $0.5$9 MHz, because the device response becomes constrained by the RC time constant. The paper models the FeCap as a leaky capacitor and states that for read pulses shorter than the RC time constant, with $\mathrm{OGLow}\approx 0.5$0 ps, the FeCap exhibits a purely resistive response. The reported solution is a new RNDR methodology using electrical read pulses down to $\mathrm{OGLow}\approx 0.5$1 ps, which are below the device RC time constant and probe a polarization-dependent leakage current rather than a standard capacitance change. The reported read energy is about $\mathrm{OGLow}\approx 0.5$2 fJ, with write time as short as $\mathrm{OGLow}\approx 0.5$3 ns and programming voltage below $\mathrm{OGLow}\approx 0.5$4 V (Baigol et al., 2 Jun 2026).

A related, but not identical, notion appears in superconducting memory. The proposed superconducting NDRO memory unit stores flux quanta in a loop and uses local feedback wiring built from JTL, SPL, and CBU cells so that a clocked read pulse is split: one copy goes to the output and the other is fed back to restore the state after readout. The reported total local reload time is $\mathrm{OGLow}\approx 0.5$5 ps, with maximum clock frequency $\mathrm{OGLow}\approx 0.5$6 GHz; the single-flux NDRO margin range is $\mathrm{OGLow}\approx 0.5$7, and the multi-flux M-NDRO margin range is $\mathrm{OGLow}\approx 0.5$8 (Ucpinar et al., 2023). The paper is explicit that this is a read-and-refresh architecture rather than an ideally non-perturbing storage primitive. In RNDR terms, it is therefore best classified as an effectively non-destructive externally clocked readout, not as passive state transparency.

5. Qubit and atom-based repetitive readout

In qubit systems, RNDR usually appears as repetitive QND or low-loss state-selective readout. A silicon electron-spin experiment demonstrated repetitive QND readout by using a neighboring electron spin in a Si/SiGe double quantum dot as an ancilla. The external field was $\mathrm{OGLow}\approx 0.5$9 T, the qubit–ancilla frequency separation was about $\mathrm{OGHigh}\approx 4$0 MHz, the induced excess exchange coupling was $\mathrm{OGHigh}\approx 4$1 MHz, and the entangling pulse duration was $\mathrm{OGHigh}\approx 4$2s. The high non-demolition fidelity, about $\mathrm{OGHigh}\approx 4$3 on average for $\mathrm{OGHigh}\approx 4$4, enabled over $\mathrm{OGHigh}\approx 4$5 readout repetitions of a single spin state, yielding an overall average measurement fidelity of up to $\mathrm{OGHigh}\approx 4$6 within $\mathrm{OGHigh}\approx 4$7 ms, and heralded preparation fidelity greater than $\mathrm{OGHigh}\approx 4$8 (Yoneda et al., 2019).

A room-temperature NV-center study analyzed repetitive QND readout of the $\mathrm{OGHigh}\approx 4$9N nuclear spin using the NV electronic spin as the optically read ancilla. The contribution was algorithmic rather than architectural: instead of a threshold on total photon counts, a neural network processed the full time trace of a repetitive readout sequence and improved fidelity by about $N$00 at the optimal repetition number and by up to $N$01 relative to the threshold method at $N$02, without additional experimental time (Liu et al., 2019). This result is notable because it treats measurement back-action as an inferable temporal signature rather than as irreducible nuisance.

Neutral-atom systems implement the same idea through low-loss fluorescence protocols. In a 2D cesium array, low-loss, non-destructive and state-selective readout on $N$03 sites achieved a factor of $N$04 suppression of the primary measurement errors, with optimized operating point $N$05 ms, survival probability $N$06, state detection fidelity $N$07, and loss-corrected detection probability $N$08 (Nikolov et al., 2023). In a separate universal neutral-atom quantum computer, non-destructive state-selective readout of cesium qubits reported state-averaged atom-loss probability $N$09, bright-dark discrimination fidelity $N$10, raw state-discrimination fidelity $N$11, and a shot cycle allowing up to $N$12 measurements per cycle, with an average shot rate of $N$13 Hz in a $N$14-qubit GHZ experiment (Radnaev et al., 2024).

The most explicitly repetitive projective-measurement neutral-atom result in the dataset uses $N$15Yb nuclear-spin qubits in optical tweezers. Under $N$16 G, near-perfect cyclicity of one nuclear spin qubit state with an optically excited state yielded bright/dark contrast of approximately $N$17 during fluorescence readout and a performance that improves as $N$18. The paper reports readout fidelity $N$19, state-preserving probability $N$20 for a single tweezer and $N$21 averaged over the array, and state-averaged readout survival $N$22 in the abstract (Huie et al., 2023). In these qubit settings, the central RNDR figures of merit are therefore not ENC or FWHM, but non-demolition fidelity, assignment fidelity, depolarization probability, and survival.

6. Applications, limiting factors, and outlook

The application space of RNDR is broad because the underlying benefit is reusable information extraction. In X-ray astronomy, SiSeRO with RNDR is presented as a route to very low-noise spectroscopic imagers for future telescopes, with region-of-interest readout, high frame rate, and sub-electron sensitivity enabling in-situ absolute calibration and improved characterization of the low-energy instrument response below $N$23 keV (Chattopadhyay et al., 2023, Chattopadhyay et al., 19 Aug 2025). In analog in-memory computing, ferroelectric RNDR addresses the tension between multilevel programming and destructive readout by making polarization-dependent state discrimination possible with repeated reads and low energy cost (Baigol et al., 2 Jun 2026). In light dark matter detection, RNDR-DEPFET is exploited in DANAE because deep sub-electron noise and high time resolution improve sensitivity to rare few-electron events (Wernicke et al., 17 Nov 2025). In neutral-atom and silicon-spin quantum computing, repeated non-destructive measurement supports mid-circuit readout, post-selection against loss, ancilla reuse, and fault-tolerant workflows (Yoneda et al., 2019, Radnaev et al., 2024).

The main limitations recur across platforms, though their microscopic origin differs. In SiSeRO, more cycles increase total readout time and can expose spectroscopy to thermal dark current unless operation is sufficiently cold; the first RNDR demonstration at about $N$24C was limited in exactly this way (Chattopadhyay et al., 2023). In Skipper-derived sensors, the cost of noise reduction is longer per-pixel readout time, motivating multi-amplifier and region-of-interest strategies such as MAS-CCD (Lapi et al., 2024, Botti et al., 14 Feb 2025). In ferroelectric FeCaps, conventional non-destructive readout collapses above about $N$25–$N$26 MHz because of the RC time constant, forcing a change of observable from capacitance to polarization-dependent leakage current (Baigol et al., 2 Jun 2026). In qubit and atom systems, the limiting factors are back-action, atom loss, or depolarization rather than amplifier noise (Nikolov et al., 2023, Huie et al., 2023).

A further extension of the RNDR idea appears in cavity-based molecular metrology. A 2026 proposal for non-destructive cavity readout of molecules uses a far-detuned high-finesse optical cavity to infer the population in a selected rotational-hyperfine state from a cavity frequency shift, with fast readout in less than $N$27 ms, variance below the standard quantum limit, and only $N$28 free-space scattered photons per molecule per readout cycle in the SrF example (Salas-Estrada et al., 1 Jun 2026). The paper argues that, for suitable cooperativity, hundreds of repeated readouts are in principle possible before the ensemble drops to $N$29 of its initial size, while practical reuse may be limited more by re-preparation efficiency than by scattering itself (Salas-Estrada et al., 1 Jun 2026). This broadens RNDR from a detector-noise technique into a general strategy for preserving scarce quantum resources during repeated interrogation.

The literature therefore presents RNDR as a unifying response to a common systems problem: conventional readout often converts information into an immediately consumptive observable, whereas many advanced applications need the state to remain available after measurement. Whether the implementation is charge shuttling between an internal gate and a summing well, transfer between two DEPFET sub-pixels, repeated sampling of a floating-gate output stage, ultrafast probing of a ferroelectric device below its RC time constant, ancilla-mediated QND spin measurement, low-loss fluorescence in atom arrays, or dispersive cavity interrogation of molecules, the technical question is the same: how to separate information extraction from state destruction strongly enough that repetition becomes useful. The current record across the cited hardware spans near-Fano-limited X-ray spectroscopy in SiSeRO, $N$30 rms in Skipper-in-CMOS, $N$31 in SiSeRO after $N$32 cycles, deep sub-electron RNDR-DEPFET operation, $N$33 fJ ferroelectric reads with $N$34 ps pulses, and repeated qubit readout with high non-demolition fidelity (Chattopadhyay et al., 2024, Chattopadhyay et al., 19 Aug 2025, Lapi et al., 2024, Wernicke et al., 17 Nov 2025, Baigol et al., 2 Jun 2026, Yoneda et al., 2019).