CMOS Image Sensors for CEvNS Detection at Nuclear Reactors

This work presents Complementary Metal-Oxide-Semiconductor (CMOS) Image Sensors (CIS) for reactor neutrino experiments targeting coherent elastic neutrino-nucleus scattering (CEvNS). The authors have developed an active-shielding strategy for physical background reduction along with an analytical model describing the impact of intrinsic detector backgrounds from dark current and readout noise at a given frame rate (fps). Assuming a conservative background of 100 DRU, and a feasible CIS-based detector with 100 g of sensitive silicon mass, 1 e− of readout noise, and 200 fps, it can detect CEvNS in less than 50 days at 16 meters from a 3.6 GW thermal reactor core. Improving the detector performance to 0.2 e− readout noise and 1000 fps substantially reduces the detection time to about 7 days. These results establish CIS technology as a promising platform for reactor monitoring and precision neutrino measurements, with thick CMOS sensors featuring parallel and non-destructive readout emerging as particularly attractive candidates for next-generation CEvNS experiments.

The paper reviews CIS technology, highlighting recent developments relevant to a CIS-based CEvNS detector. Unlike CCDs, whose serial pixel readout results in long readout times, CIS architectures enable parallel readout of the pixel array, supporting frame rates of up to 1000 frames per second (fps). Advances driven by the microelectronics industry have reduced the readout noise of CMOS sensors to below one electron. More recently, skipper-style readout has been demonstrated in CMOS processes, enabling repeated non-destructive measurements and, consequently, an arbitrary reduction of the readout noise.

The authors propose an active shielding strategy exploiting the counting Poisson statistics of cosmic muons together with a sufficiently high frame rate. Frames acquired in coincidence with a trigger from the muon detector are rejected. Assuming a muon rate of 200 Hz, the expected number of detected muons per frame acquisition time is λ = 200 Hz / fps. For example, at 1800 fps, 90% of the frames are accepted, and the acquisition time of the accepted frames is the detector live time. This approach allows a live time exceeding 90%.

The two main sources of intrinsic detector background (IDB) of a CIS are thermal generation of electron-hole pairs (dark current) and the amplifiers' inherent electronic noise (readout noise). The paper presents an analytical model for estimating event rates in pixelated detectors, based on the counting of geometrical shapes known as polykings. This model computes the probability of forming events with arbitrary numbers of connected pixels, and it has been applied to predict background events from IDB that have a number of pixels in the same range as the signal events (at most four pixels).

The expected CEvNS signal in silicon targets is calculated, and the achievable detection significance is evaluated for representative nuclear reactor scenarios. Two representative experimental configurations are considered: H1 = 0.014 GWth/m2, corresponding to near-core, high-flux setups such as ATUCHA, CONUS, NUCLEUS, and PROSPECT; and H2 = 0.004 GWth/m2, representing configurations with reduced neutrino flux, similar to CONNIE, NEON, and TEXONO. The expected signal rate dR/dE for each configuration is shown in Fig. 4.

The paper assesses the discovery potential of a CIS-based experiment by treating the measurement as a single-bin counting experiment in the energy range [Emin, Emax]. The median discovery significance Z for a Poisson counting experiment with known background is obtained from the profile-likelihood ratio in the asymptotic (Asimov) approximation. A two-dimensional scan over the lower and upper integration limits is performed to determine the spectrum energy range that maximizes the significance for CEvNS detection. For the case of H1, a significance of 3σ is achieved for the window [0.0576, 0.167] keVee.

Considering a feasible CIS-based detector with the characteristics listed in Table I (readout noise of 1 e−, dark current of 1.5 e−/pix/s, silicon thickness of 725 µm, 1 Mpix, pixel size of 15 × 15 µm2, 1000 fps, 256 CISs, total silicon mass of 97 g), the IDB contribution is calculated. The detector threshold must be over 0.064 keV to suppress the IDB contribution. For the scenario H1, to reach CEvNS an observation time of 17 days is necessary. In contrast, for H2 the observation time is equal to 150 days. The inefficiency due to the live time given by the active shielding is already included in the calculation and increases the observation time by 20%.

The analytical framework is applied to analyze other detector configurations. Table II summarizes the results for the H1 scenario assuming a physical background of 100 DRU and 100 g of detector mass. In all configurations, 3σ of significance is reached in less than 50 days. In the case of H2, about an order of magnitude larger than H1 due to its lower signal expectation. This demonstrates that reactor CEvNS detection is already feasible with 1 e− noise CIS in a short time scale, even while progress towards sub-electron performance would substantially reduce the time to the CEvNS detection.

Two main applications are identified: one primarily oriented toward technological developments and applied neutrino detection (readout noise of 1 e− and 200 fps already enables reactor monitoring applications, detector characterization, and the validation of low-energy response models in silicon-based devices), and another focused on fundamental physics searches (requiring sub-electron readout noise to improve the signal-to-background ratio and extend the sensitivity of searches for non-standard neutrino interactions, light mediators, or other exotic scenarios like millicharged particles). The paper concludes that CIS technology provides a realistic pathway toward this goal, with additional multiple non-destructive sampling capability allowing the readout noise to be reduced below the single-electron level while maintaining a high frame rate. Such advances would open the door to precision CEvNS measurements and competitive probes of new physics in the low-momentum-transfer regime.

Improvements to AI Systems:

1. Physics-Aware Event Reconstruction: Implement a novel AI architecture that integrates the paper’s analytical polyking model for pixel-cluster counting. The AI can directly classify and reject background events (dark current, readout noise) by predicting the probability of multi-pixel clusters, reducing false positives in low-energy neutrino detection.

2. Adaptive Frame-Rate Optimization: Train a reinforcement learning agent to dynamically adjust CIS frame rates (e.g., between 200–1000 fps) based on real-time muon trigger rates and dark current fluctuations. This maximizes live time (up to 90%+ as shown) while minimizing intrinsic detector background, improving signal-to-noise ratio autonomously.

3. Noise-Robust Threshold Prediction: Develop a deep learning model that predicts optimal energy thresholds (e.g., >0.064 keVee) from simulated detector parameters (readout noise, dark current, pixel size). This enables automatic calibration for new CIS configurations, reducing manual tuning and accelerating detector deployment.

4. Asimov-Based Significance Forecasting: Create a neural network that takes detector specs (mass, fps, noise, distance from reactor) and outputs median discovery significance (Z) using the profile-likelihood ratio approximation. This allows rapid feasibility studies for new reactor sites or detector upgrades without full Monte Carlo simulations.

5. Multi-Scale Signal Extraction: Implement a convolutional neural network (CNN) trained on simulated CEvNS signals (1–4 pixel events) and background patterns. The CNN can separate signal from dark-current-induced clusters at frame rates up to 1000 fps, enabling real-time event filtering for reactor monitoring.

Capabilities of the Improved AI System:

• Autonomous Detector Control: The AI can self-tune frame rate, threshold, and shielding logic to achieve 3σ CEvNS detection in <50 days (or <7 days with sub-electron noise) without human intervention, adapting to changing reactor power or environmental backgrounds.

• Rapid Experiment Design: Researchers can input target sensitivity (e.g., 5σ) and receive optimized detector parameters (mass, fps, noise, distance) within seconds, reducing design cycles from months to minutes.

• Real-Time Anomaly Detection: The system can distinguish neutrino-induced events from cosmic muon coincidences and thermal noise in real time, enabling continuous reactor monitoring for non-proliferation or power output verification.

• New Physics Sensitivity: By leveraging sub-electron readout predictions, the AI can flag rare events consistent with non-standard interactions, light mediators, or millicharged particles, extending searches beyond standard CEvNS.

• Cross-Platform Transferability: The trained models can be applied to other pixelated detectors (e.g., CCDs, Skipper-CCDs) with minimal retraining, generalizing the active-shielding and background-rejection strategies to broader low-energy physics experiments.

In this work we present Complementary Metal-Oxide-Semiconductor (CMOS) Image Sensors (CIS) for reactor neutrino experiments targeting coherent elastic neutrino-nucleus scattering (CEvNS). We have developed an active-shielding strategy for physical background reduction along with an analytical model describing the impact of intrinsic detector backgrounds from dark current and readout noise at a given frame rate (fps). Assuming a conservative background of 100 DRU, and a feasible CIS-based detector with about 100g of sensitive silicon mass, 1 e- of readout noise, and 200fps, it can detect CEvNS in less than 50 days at 16 meters from a 3.6 GW thermal reactor core. Improving the detector performance to 0.2 e- readout noise and 1000fps substantially reduces the detection time to about 7 days. These results establish CIS technology as a promising platform for reactor monitoring and precision neutrino measurements, with thick CMOS sensors featuring parallel and non-destructive readout emerging as particularly attractive candidates for next-generation CEvNS experiments.