The cryogenic and purification systems to maintain a stable operation of the LXe detector have been developed in the particle physics field and many have become standard technology. For example, the pulse tube refrigerator used in the XENON1T dark matter detector (Aprile et al., 2017) provides 250 W cooling power at the LXe temperature around - 9 6 o C . Unlike the dark matter detector, the LXePET system does not require low background materials but the liquid xenon itself needs to be constantly circulated with a pump and purified through a getter, such as the SAES PS4-MT50-R used in XENON1T, to remove electronegative impurities to allow free drift of ionization electrons in the liquid. The detailed engineering design of these systems for the LXePET is out of the scope of this paper but we envision smaller scale systems will be needed to operate the LXePET detector, which in our design contains less than 100-kg of liquid xenon, compared to the ton-scale dark matter detector.

A design of the central part of our proposed PET imaging system with individual LXe detectors is shown in Figure 1A. The imaging system contains concentric rings of individual detectors (PET detectors) of about 5 × 5 × 5   c m 3 , filled with liquid xenon as the gamma-ray detection medium. The LXe PET detector will detect both the prompt scintillation (S1) and delayed ionization signals through field-enhanced electroluminescence (S2) to enable high energy and position resolutions.

Field-enhanced electroluminescence in LXe was first demonstrated by Aprile et al. (2014) in a single-wire test chamber. Recent development has shown that stable operation on a large kg target can be performed in a single wire proportional scintillation counter (Qi et al., 2023) and time projection chamber (TPC) with thin wire anode (Tönnies et al., 2024). Here we follow the concept from (Breskin, 2022), which proposes several novel ways including the use of a micro-strip anode plate to generate electroluminescence in LXe, to design our PET detector, with its principle shown in Figure 1B. When a particle, such as the 511 keV gamma ray from the positron radioactive source, enters the liquid xenon medium, it deposits energy and produces scintillation photons (S1) and ionization electrons, which drift towards the anode plate under an electric field. A strong field near the strip on the anode converts these electrons into electroluminescence (S2). The same photo-sensor array, located right above the anode plane, detects both S1 and S2 to reconstruct the energy and position of the energy deposition.

We use SOLIDWORKS, a 3D CAD software, to create mechanical prototype design for the LXe PET detectors (Figure 2). This software helps with measurements and the general design of the PET detectors with Teflon walls containing the liquid xenon target in a 5 × 5 × 5   c m 3 detection volume. The size of the PET detector is chosen to efficiently detect the 511-keV gamma rays interacting in LXe while minimizing the PET detector numbers for system complication and cost. The highly reflective Teflon walls improve light detection (see Section 3.1), which is essential in determining the energy resolution. The detector is instrumented with electrodes and shaping rings to optimize the detection of ionization electrons (see Section 3.2. An array of photosensors detect both the primary scintillation and electroluminescence from ionization. The electroluminescence has similar spectrum range to the scintillation light from liquid xenon. The design uses 3 × 3 Hamamatsu S13371-6050CQ-02 silicon photomultipliers (SiPMs), each having a photon detection area about 144 m m 2 . The SiPMs feature both a compact mechanical structure and are capable of detecting the 175 nm vacuum ultra-violate (VUV) photons from LXe with a photon detection efficiency about 24%. It is worth to note that the MEG experiment used this type of SiPMs and observed UV-induced degradation Ieki et al. (2023). Given the EL light exposure rate in our PET design, long-term SiPM stability should be further investigated in an experimental setting. One notable instance of design oversight pertains to the distance between the anode and photo sensors, a critical parameter to position accuracy. Machine learning algorithms are developed (see Section 4) to investigate and optimize such a distance.

To validate the design, we use GEANT4 (Agostinelli et al., 2003), an open-source particle simulation software, to simulate particle trajectories and collect data, aiding in the design process. We implement our design in GEANT4 to trace light generated in the detector and detected by the SiPMs, comparing the light collection efficiency for different reflectivity of the Teflon wall. Additionally, we utilize COMSOL Multiphysics COMSOL Inc. (2024), a simulation software that aids in studying electromagnetism for generating electric fields. COMSOL enables us to examine electrode configurations and electric field uniformity which is crucial for ionization electron detection efficiency.

Specific pieces that went through multiple design iterations are shown. For example, the thickness of Teflon was important as we needed to factor in space for shaping rings as well as the rigidness of the cathode anode. Electric field simulations ultimately determined that this space should be created as uniform an electric field as possible. The stainless-steel mesh was important to allow for as much light to pass through as possible. Finally, we compared the pros and cons of adding SiPM sensors in a 2 × 2 versus 3 × 3 . While a 2 × 2 SiPM sensor is more cost-efficient, a 3 × 3 SiPM sensor layout will enlarge the detector size to 5 by 5 cm, allowing the detector to have higher light collection efficiency and electric field uniformity.

In our proposed LXe PET detector, the gamma rays depositing energy in the monolithic LXe target produce a prompt scintillation signal (S1), followed by a delayed ionization-induced electroluminescence signal (S2). The time difference between the S1 and S2 provides the depth-of-interaction (DOI) information with a sub-mm spatial resolution (Tönnies et al., 2024). The S2 light pattern, collected by the 3 × 3 photosensor array above the anode, reconstructs the horizontal (X&Y) positions. We investigate two methods of reconstructing the gamma-ray’s XY position in the LXe PET detector.

Both methods consist of training machine learning models. The first method constructs a multilayer perceptron (MLP) model (Taud and Mas, 2018), while the second constructs a model using XGBoost (Chen and Guestrin, 2016). To train each model, we take the nine numbers corresponding to the number of photons each SiPM detected, feed it to the model, and then evaluate model performance based on how close the reconstructed gammy ray position was to the original position.

To gather the data needed to perform and evaluate these methods, the PET detector was constructed in GEANT4, and the generation of the photons from a single point due to the gamma-ray interacting with the PET detector was simulated. Due to the symmetrical nature of the PET detector, points in a single quadrant of the entire sensitive target, with a separation of 0.2 mm between each point, were equally spaced and generated in the PET detector. The generated data was then reflected across the axes to get the data for the other three quadrants. We vary the anode distance relative to the SiPM array in the PET detector and find the best position reconstruction results that can be achieved is at an anode distance about 10 mm below the SiPM array. In the following, we present results for an anode distance 10 mm below the SiPM array.

We define “accuracy” as the relative distance between the reconstructed position and the original position. The reconstructed position accuracy at different X&Y positions is shown in Figure 8 for the two methods. Both methods achieve an average accuracy value around zero, indicating no systematic bias. The MLP model shows a wider distribution (Figure 9) than the XGBoost model. The MLP method performs better when the gamma-ray position is near the center of the PET detector, and poorly near the edges and corners. The XGBoost model performs well in the entire target volume.

Plotting the difference in positions, for both X&Y, between the original and reconstructed positions reveals a distribution centered around zero for the MLP method, with a wider base corresponding to the edges and corners of the cell, as shown in Figure 9. For the XGBoost method, there is Gaussian-like distribution centered at zero. Fitting the distribution with a Gaussian function reveals a FWHM resolution of 0.56 mm.

From Figure 8, we can see poorer position reconstruction near the edge of the detector. This is due to the increased inaccuracies of the models near the edges. We performed further study by removing points within 5 mm of the edge of the PET detector, and then the models were trained and tested on all the remaining points. Removing the 5-mm from the edge of detector reduces the sensitive target to about 67% of the total volume. While the MLP method’s shows slightly better performance, it still shows a wider base (Figure 10, left). The XGBoost model achieves an improved FWHM position resolution to 0.34 mm in the central volume.

Although there are various methods for using machine learning to reconstruct positions and optimize target volume selection, the sub-millimeter position sensitivity of the proposed LXe PET design offers the potential to enhance the overall position sensitivity of a complete PET system. This improvement could lead to a reduction in the radioactive dose required for patients.

Our study highlights the potential of liquid xenon (LXe)-based detectors to significantly enhance the performance of Positron Emission Tomography (PET) imaging systems. By combining scintillation and ionization energy detection, the proposed design overcomes the limitations of traditional scintillation crystal-based PET systems. Based on our simulations, the design achieves an energy resolution of approximately 2.1% (FWHM), representing a five-fold improvement over current PET technology. Additionally, sub-millimeter position resolution can be attained for gamma-ray interaction points using machine learning models. The LXePET will use mainly the single-scatter (SS) photo-absorbed 511-keV gamma events to reconstruct the position of the radioactive source. The gamma rays can make Compton multiple scatterings on the cryostat and the liquid xenon target itself. Selecting SS from MS events thus will be the key for enhanced performance. We ran a Geant4 simulation of 511-keV gamma rays interacting in the LXePET and used the simulated spatial resolution (0.56 mm FWHM) for identifying the SS/MS events. Figure 11 shows the SS and MS events identified in the LXe target. In the 2.1% (FWHM) window around the 511-keV peak. The SS event accounts for about 44% of the events detected. While the pure SS photo-absorbed events will improve the time-of-flight reconstruction precision, the first interaction of fully contained multiple scattering events can be used for time-of-flight and energy reconstruction. The multiple Compton scattering events can be used as well to reconstruct source positions, as used in the XEMIS2 detector for medical imaging Manzano et al. (2018).

One consideration of using both scintillation and ionization signals for LXePET is the event pile-up. At the design field of 1 kV/cm, the electron drift speed can reach about 2 mm/ μ s Albert et al. (2017), giving a maximum drift time ( d t max ) about 25 μ s across the 5-cm distance from the cathode to anode. The probability of two events piling up in the same d t max window can be estimated by multiplying the event rate. The pile-up probability is ∼ 13% for an event rate of 5 kHz. Such an event rate would be reasonable to operate the LXePET detector in both scintillation and ionization mode to identify enough good events without pile-ups. Event selections, such as selecting events in the 511-keV energy window and pairing S1 and S2 events with their energy information to identify separate events, will further suppress the pile-up probability. Assuming the distance between the single LXePET cell window ( 5 × 5 cm 2 ) and the center of the ring (Figure 1A) of 20-cm, the corresponding radiation dosage is estimated to be around 1 MBq to produce 5 kHz events in the cell. The image reconstruction capability of the whole LXePET detector, taking into account the low radiation dosage, SS/MS event reconstruction and background events, will be investigated in a future study.

In our design, the 5-cm thick liquid xenon (LXe) detector, with a density of 2.89 g/cm³; and an attenuation length of 3.7 cm (Gomez-Cadenas et al., 2016) for 511-keV gamma rays, can achieve an estimated detection efficiency of 74%. This is comparable to similar thickness NaI(Tl) detector but is slightly lower than LYSO/BGO crystals with more than 90% efficiency for typical PET detector thicknesses of 2-cm. Increasing the thickness of the LXe detector will increase the pileup rate but also the cost of material. However, the high position and energy resolutions effectively reduces background events in the 511-keV gamma-ray energy window, thereby lowering the required patient exposure time and radiation dose. The enhanced position resolution further improves the accuracy of reconstructed positions within the patient’s body. The efficiency of our monolithic LXe target design not only provides more precise positioning and accuracy but also simplifies the mechanics and reduces costs if implemented in medical imaging systems.

While our simulated results are promising, this work only presents the potential of a LXeTPC for PET applications. The practical implementation of a LXe PET detector still needs to be demonstrated. Although the detection of liquid xenon scintillation light with SiPMs has been successfully achieved in several particle physics experiments, the detection of ionization-induced electroluminescence in liquid xenon has only been demonstrated in small-scale, table-top prototypes (Qi et al., 2023; Martinez-Lema et al., 2024; Qi et al., 2025). The reliability and performance of the proposed LXe PET detector must be validated and optimized for real-world applications, with experimental efforts required to address any unforeseen challenges that may arise during the transition from simulation to laboratory testing.