TAO: Taishan Antineutrino Observatory

• TAO is a reactor antineutrino detector employing gadolinium-doped liquid scintillator to achieve better than 2% energy resolution at 1 MeV.
• It integrates dense silicon photomultiplier arrays, advanced calibration systems, and low-background engineering to deliver high-statistics spectral measurements.
• TAO serves as a near reference for JUNO, resolving reactor spectrum anomalies and validating technologies for reactor monitoring and neutrino oscillation studies.

TAO, in the reactor-neutrino context, denotes the Taishan Antineutrino Observatory, a tonne-scale gadolinium-doped liquid scintillator satellite experiment of the Jiangmen Underground Neutrino Observatory (JUNO). It is designed to measure the reactor antineutrino energy spectrum with better than 2% at 1 MeV energy resolution, and its instrumentation program couples ultra-high-light-yield calorimetry, dense silicon photosensor coverage, detailed calibration, low-background engineering, and precision event reconstruction (Shi et al., 8 Aug 2025). Within the broader JUNO program, TAO serves as a near, high-statistics spectral reference and as a technology demonstrator for reactor monitoring and related detector R&D (Steiger, 2022).

1. Scientific rationale and role within JUNO

TAO is a satellite detector for JUNO, located near the Taishan nuclear power plant, and is intended to provide a high-precision measurement of the reactor electron antineutrino spectrum before the oscillation effects relevant to JUNO’s medium-baseline program become dominant (Steiger, 2022). The experiment is described as having several linked objectives: high-precision measurement of the reactor antineutrino spectrum at a very short baseline, provision of a reference spectrum for JUNO and other future experiments, resolution of reactor-spectrum discrepancies such as the 5-MeV “bump” and the reactor antineutrino anomaly, searches for new physics including light sterile neutrinos, tests of nuclear database predictions, and technological validation for reactor monitoring and nuclear safeguard applications (Shi et al., 8 Aug 2025).

The expected neutrino detection rate is about 2000 per day, approximately 30 times the rate in the JUNO main detector (Steiger, 2022). This high rate, combined with superior calorimetric performance, is central to TAO’s function as a precision spectral instrument rather than merely a normalization monitor. In the literature on TAO–JUNO joint analyses, this role is formalized as the provision of a directly measured, high-resolution reactor reference spectrum that can be propagated to the far detector through a controlled response mapping (Capozzi et al., 2020).

2. Detector architecture and instrumentation

The TAO central detector is a 2.8 tonnes volume of Gadolinium-doped Liquid Scintillator (GdLS) contained in a 1.8 m-diameter acrylic vessel (Shi et al., 8 Aug 2025). Photosensor coverage is unusually dense: one detector description gives 4024 SiPM tiles with 2 readout channels each, for 8048 channels total and more than 93.5% optical surface coverage, while another summarizes the system as an array of approximately 4,000 silicon photomultipliers with about 95% optical coverage (Shi et al., 8 Aug 2025, Steiger, 2022). The design objective is near-complete photon collection compatible with better than 2% at 1 MeV energy resolution (Shi et al., 8 Aug 2025).

TAO’s calorimetric concept combines high PDE SiPMs with low-temperature operation. The detector is described as operating at $-50^\circ\mathrm{C}$ to reduce SiPM dark noise and increase light yield, and the combined cold-LS and cold-SiPM configuration is reported to provide a 4.5-fold increase in photoelectron yield relative to conventional setups (Steiger, 2022). The neutrino interaction channel is inverse beta decay,

$\bar{\nu}_e + p \rightarrow e^+ + n ,$

with neutron capture on gadolinium providing a delayed tag through an approximately 8 MeV gamma cascade and a capture time of about 28 \,\mu\mathrm{s} (Steiger, 2022).

Shielding and veto instrumentation are layered. The experiment uses passive shielding with water, high-density polyethylene, polyurethane, and lead, together with active systems including a plastic scintillator array and a water Cherenkov detector (Steiger, 2022). These systems are not ancillary: at TAO’s shallow site, external and cosmogenic backgrounds are a first-order design constraint rather than a secondary correction.

3. Calibration and control of detector response

TAO’s calibration strategy is built around two source-deployment systems: the Automated Calibration Unit (ACU) and the Cable Loop System (CLS) (Xu et al., 2022). The ACU supports deployment along the detector axis and includes a UV LED source, a $^{68}$Ge positron source, and a combined multi-gamma and neutron source. The CLS provides off-axis source deployment and is used in particular with a $^{137}$Cs source to map spatial response away from the central line (Xu et al., 2022).

The deployed source suite covers the prompt-energy region relevant to reactor antineutrino detection. The calibration summary explicitly lists $^{137}$Cs, $^{54}$Mn, $^{60}$Co, $^{40}$K, $^{68}$Ge, $^{241}$Am–$\bar{\nu}_e + p \rightarrow e^+ + n ,$0C, neutron capture on hydrogen, and cosmic-muon-induced $\bar{\nu}_e + p \rightarrow e^+ + n ,$1B beta decay (Xu et al., 2022). The associated non-linearity model for electrons and positrons is written as

$\bar{\nu}_e + p \rightarrow e^+ + n ,$2

with quenching and Cherenkov contributions treated separately (Xu et al., 2022).

The calibration paper reports that the non-linear energy response can be controlled within 0.6% using radioactive-source calibration points, and that it can be further improved through the inclusion of $\bar{\nu}_e + p \rightarrow e^+ + n ,$3B decay signals produced by cosmic muons (Xu et al., 2022). Through non-uniformity calibration, the residual non-uniformity is less than 0.2%, while the resulting energy resolution degradation and energy bias can be controlled within 0.05% and 0.3%, respectively (Xu et al., 2022). The same system includes a dedicated ultraviolet LED subsystem for monitoring SiPM gain, timing, and quantum efficiency (Xu et al., 2022).

4. Background environment and mitigation

TAO is a shallow-overburden experiment, and the overburden is described as only about 10 meter-water-equivalent (Li et al., 2022). This induces a large cosmogenic-neutron burden. In the detector-optimization study, the cosmogenic neutron background-to-signal ratio was initially estimated to be about 10%, with detailed Monte Carlo indicating substantial contributions from fast neutrons and especially double-neutron topologies (Li et al., 2022).

The primary mitigation measures identified in simulation are doping gadolinium in the buffer liquid, adding a polyethylene layer above the bottom lead shield, and optimization of the veto strategy (Li et al., 2022). With these changes, the study reports that the cosmogenic neutron background-to-signal ratio can be reduced to about 2%, and may be further suppressed with pulse shape discrimination (Li et al., 2022). A more detailed summary gives a final post-optimization rate of 44/day cosmogenic neutron backgrounds, an IBD signal rate of 1838/day, and a detector dead time of 8.1% under the optimized veto (Li et al., 2022).

Independent ambient-neutron measurements at the TAO site used a Bonner sphere spectrometer and iterative Maximum-Likelihood Expectation-Maximization (MLEM) unfolding (Li et al., 2022). The total neutron fluence rate was measured to be $\bar{\nu}_e + p \rightarrow e^+ + n ,$4, higher than the Geant4 expectation of $\bar{\nu}_e + p \rightarrow e^+ + n ,$5 (Li et al., 2022). The study reports that the measured fluence rate and energy spectrum below 20 MeV can be reproduced by simulation, whereas a significant discrepancy remains above 20 MeV and requires further investigation (Li et al., 2022). This result is significant because high-energy neutrons are the most difficult to moderate and shield.

5. Vertex reconstruction, visualization, and data acquisition

Precise vertex reconstruction is explicitly identified as necessary for fiducialization, uniformity correction, and systematic-uncertainty control, with a stated TAO requirement that vertex resolution and bias both be better than 5 cm (Shi et al., 8 Aug 2025). Two complementary reconstruction methods have been developed: the charge center algorithm (CCA) and a deep learning algorithm (DLA) (Shi et al., 8 Aug 2025).

The CCA reconstructs the event position as a charge-weighted centroid,

$\bar{\nu}_e + p \rightarrow e^+ + n ,$6

followed by detector-specific corrections for the dual-opening geometry, nonlinear radial calibration, and electronic noise, especially dark noise (Shi et al., 8 Aug 2025). The DLA uses projected 2D SiPM maps of charge and first-hit time, with two CNN architectures, VGG-T and ResNet-T; the reported training configuration uses 240,000 training events, 60,000 validation events, and 20,000 test events per energy/type, with Smooth L1 loss plus boundary penalties (Shi et al., 8 Aug 2025).

At 1 MeV, the optimized methods achieve the following performance (Shi et al., 8 Aug 2025):

Method	Radial performance	Angular performance
CCA (optimized)	resolution $\bar{\nu}_e + p \rightarrow e^+ + n ,$7 mm, bias $\bar{\nu}_e + p \rightarrow e^+ + n ,$8 mm	$\bar{\nu}_e + p \rightarrow e^+ + n ,$9, $^{68}$0
DLA (ResNet-T)	resolution $^{68}$1 mm, bias $^{68}$2 mm	$^{68}$3, $^{68}$4

The DLA also reports angular biases below $^{68}$5 in both $^{68}$6 and $^{68}$7 at 1 MeV (Shi et al., 8 Aug 2025). The comparison indicates that deep architectures, especially ResNet-T, are superior in radial resolution and bias, while the CCA remains a simple and physically transparent baseline.

Operationally, TAO is supported by a ROOT-based detector geometry and event visualization system embedded in the offline framework and built on SNiPER, ROOT, and EVE (Liao et al., 2024). It visualizes detector geometry, simulation, calibration, and reconstruction events, provides 3D and 2D projections including Aitoff views, and is intended for detector commissioning, reconstruction tuning, physics analysis, and reactor monitoring (Liao et al., 2024). Data taking is handled by a distributed DAQ system divided into a data flow system and online software, with the onsite storage bandwidth limited to less than 100 Mb/s (Zhang et al., 2024). The DAQ is reported to have been deployed and successfully applied to detector and electronics prototype integration tests (Zhang et al., 2024).

6. Spectral transfer to JUNO and impact on oscillation analyses

A dedicated TAO–JUNO study formulates the near-to-far connection as a response mapping from the TAO visible-energy spectrum to the JUNO visible-energy spectrum (Capozzi et al., 2020). In the absence of oscillations, the mapping is stated to be exact:

$^{68}$8

where the additional Gaussian smearing $^{68}$9 encodes the difference in quadrature between the JUNO and TAO resolutions (Capozzi et al., 2020). In the physical case with oscillations, the paper introduces an effective oscillation probability in visible energy and reports agreement with the fully convolved oscillated spectrum at the $^{137}$0 level (Capozzi et al., 2020).

This mapping matters because reactor spectra may contain fine structure from summation calculations, and both energy resolution and nucleon recoil suppress observable substructure (Capozzi et al., 2020). The same study concludes, through a chi-squared analysis of bundles of variant spectra generated by changing nuclear input uncertainties, that these variants induce only a small reduction in JUNO’s mass-ordering sensitivity, especially when TAO constraints are included (Capozzi et al., 2020). In this sense, TAO is not merely a normalization anchor; it is a spectral-control instrument that compresses reactor-model uncertainty before the far-detector fit.

The broader implication is that TAO provides JUNO with a directly measured, high-resolution reactor reference spectrum against which oscillated spectra can be compared. This suggests a division of labor in which TAO constrains source and detector-response systematics near the core, while JUNO exploits the long baseline for precision oscillometry and mass-ordering sensitivity (Steiger, 2022).

7. Other scientific uses of “TAO” in arXiv literature

The acronym TAO is also used for several unrelated research systems and datasets. In computer architecture, “TAO: Re-Thinking DL-based Microarchitecture Simulation” is a deep-learning simulation framework that uses functional traces, self-attention, and microarchitecture-agnostic embeddings to reduce overall training and simulation time by 18.06x over prior DL-based efforts (Pandey et al., 2024). In computer vision, “Track Any Anomalous Object (TAO)” denotes a granular video anomaly detection pipeline that combines object-centric anomaly scoring, robust temporal filtering, and SAM2-based segmentation (Huang et al., 5 Jun 2025), while “TAO: A Large-Scale Benchmark for Tracking Any Object” designates a tracking dataset with 2,907 high-resolution videos and 833 categories (Dave et al., 2020).

In approximate nearest-neighbor search, “Tao” is a framework for terminating ANN queries adaptively using only static features via predicted local intrinsic dimension (LID), with reported speedups of up to 2.69x over AdaptNN at matched accuracy targets (Yang et al., 2021). In electronic-structure theory, TAO-DFT denotes thermally-assisted-occupation density functional theory, and hybrid TAO-DFT schemes incorporate exact exchange to improve treatment of nonlocal exchange effects and strong static correlation (Chai, 2016). In time-domain astronomy, Deep-TAO is an open dataset of 1,249,079 annotated images for transient astronomical object classification (Suárez-Pérez et al., 20 Mar 2025). These usages are terminologically independent of the Taishan Antineutrino Observatory.