A team of scientists from MIT and Lawrence Berkeley National Laboratory (LBNL), led by Ryotaro Okabe, has developed a computational model to design a simple and streamlined radiation sensor setup that can determine the direction of radiation sources. By moving the sensor and taking multiple readings, this technology can provide cross-referenced bearings to provide accurate location.
Radiation detectors have already been developed to determine the location and intensity of radiation sources, but the complex mechanisms of radiation-matter interaction have traditionally made high-performance, low-cost radiation mapping difficult.
Radiation is typically detected using semiconductor materials such as cadmium zinc telluride (CZT), which generate an electric current when struck by radiation such as gamma rays. However, radiation easily penetrates matter, making it difficult to determine the direction from which such a signal is emitted. The large penetration depth of the radiation reduces the angular sensitivity of the detector and limits most radiation detection efforts to focus on counting and spectral acquisition rather than directional information. They cite the Geiger counter as an example, which simply gives a clicking sensation when it receives radiation, but provides no information about the energy or type of radiation. Therefore, to find the source, the detector must be moved to find the maximum intensity. Therefore, this process requires the user to be close to the radiation source, potentially increasing the risk of exposure.
The challenge of obtaining directional radiation information makes source localization and radiation source location even more difficult. To provide directional information from a fixed device without getting too close, researchers use an array of detector grids along with a mask. This mask produces different patterns on the array depending on the direction of the light source. For example, the High Efficiency Multimode Imager (HEMI) consists of two layers of CZT detectors. The first layer has randomly placed apertures, and the second layer is a conventional coplanar detector grid. This system requires the incident beam to enter from only a limited range of angles in order to pass through the aperture in the first layer and interact with the second layer. Algorithms interpret the varying timing and strength of signals received by individual detectors or pixels to provide direction, but this approach often complicates detector design. Generally, such detector arrays capable of sensing the direction of the radiation source, comprising a 10×10 array and at least 100 pixels, are large and expensive. Not only are the individual sensing elements more expensive, but all the interconnections that carry the information from those pixels are also much more complex.
Traditional reconstruction algorithms also require that all incoming beams fall within the field of view. Accuracy is affected if the radiation comes from a different direction, especially for near-field radiation. Moreover, this type of system is usually only conditionally capable of detecting multiple sources if the sources originate from different isotopes and can be distinguished by energy. In this scenario, multi-source detections can be reduced to single-source detections by considering only the number of events within an energy range. However, in real-world applications, it is not always possible to distinguish between different sources within the energy spectrum. Another approach uses detectors separated by pad shield material. Radiation sources from different directions and distances can result in different intensity distribution patterns on the detector array. However, it is difficult to extract information from detector data using these traditional methods, as misalignment of the detector and shielding material and manufacturing errors lead to inaccurate models. Therefore, although there are versions of simplified arrays for radiation detection, many are only effective when the radiation comes from a single local source and can be confounded by multiple or dispersed sources.
Alternative approaches to detection
As an alternative, a joint team at MIT and LBNL has developed a radiation mapping framework using detector pixels inspired by the video game Tetris. said Okabe, MIT professors Minda Li and Benoît Fauget, senior researcher Lin Wen Hu, and principal investigator Gordon Kose, according to the research team’s recent paper published in Nature Communications. Graduate student Shangjie Xue. LBNL research scientist Jayson Vavrek and many other researchers at both institutions have found that using just four pixels arranged in the tetromino shape of the “Tetris” game figures can nearly match the accuracy of larger, more expensive systems. The Tetrominoes family of shapes consists of four squares. The authors’ radiation detection framework uses a minimal number of detectors, combines Tetris-shaped detectors with interpixel padding, and deep neural network-based detector readout analysis. Intentionally includes padding material between pixels to increase contrast. One of the big advances in making the system work is placing insulation between the pixels to increase the contrast between radiation readings entering the detector from different directions. Therefore, the contrast between pixels is not only generated by the angle of incidence, but also enhanced by the pad layer, which is a good absorption layer of radiation. The pad material between the pixels is chosen to be 1 mm of lead, which is thick enough to create contrast and has very low absorption of photons in the gamma range. Leads perform the same function as the more elaborate shadow masks used in larger traditional systems.
The detector pixels are each made of CZT with a size of 1 cm², slightly larger than the crystals used in current detectors, but still much smaller than the 5 meter distance from the source to the detector. The detector grid is placed horizontally in the plane rather than facing perpendicular to the source.
The key to this detection method is the computer’s appropriate reconstruction of the angle of arrival of the rays, based on the time each sensor detected the signal and the relative strength detected by each sensor, reconstructed through an AI-guided study of the simulated system. Machine learning algorithms have been implemented to analyze detector readings and have demonstrated great promise in reducing the need for detector pixel count, thereby reducing manufacturing and implementation costs.
The researchers tried different configurations of four pixels to evaluate the predictive accuracy of a detector consisting of a 2 × 2 square grid and four detector configurations: S-, J-, and T-shaped tetrominoes. The I-shaped Tetris detector array is not shown as it does not perform well for directional mapping.
The research team found that a less symmetrical arrangement yielded more useful information from a smaller array. The S-shape detector operated on a 2 × 2 square, J-shape, and T-shape following the minimum prediction. The T-shaped Tetris is the least accurate, but all four types of detectors work well enough to determine the direction of the radiation source with an accuracy of about 1°.
Throughout the study, the authors assume that the incident beam energy is 0.5 MeV of γ-rays, which they suggest is a realistic energy from pair production and comparable to many γ-decay energy levels.
Evolving detection AI
Using a neural network trained on Monte Carlo (MC) simulation data, high-resolution direction prediction can be achieved with a detector of as few as 4 pixels, demonstrating that this approach has the potential to achieve higher resolution than more complex alternatives. The authors suggest that experimental validation could further demonstrate the ability of the machine learning approach to localize radiation sources in real-world scenarios.
The authors point out that in real-world radiation mapping applications, it is highly desirable to go beyond directional information to determine the exact location of a radiation source. They propose a method based on maximum a posteriori (MAP) estimation that generates the distribution of radiation through simple detector movements. The detector readout is simulated by the MC by specifying the initial position and orientation of the detector before it starts its circular motion. During detector operation, a predicted source direction is calculated based on the model, and then the source position is estimated via MAP based on a set of detector direction data estimated by a neural network at different detector positions. In the ideal case of a single radiation source, only two positions are sufficient to locate the radiation source, but circular motion and MAP are implemented for more complex radiation profile mapping.
Therefore, the use of moving detectors in MAP has furthered this development and successfully localized the radiation source. Subsequent field tests with a simple detector in a single-blind field test at Berkeley Lab using a real cesium radiation source verified the ability of the MAP method to determine both the direction of the radiation source and the distance to the source with high accuracy. Although Vavrek led this element of the study, the MIT researchers did not know the location of the radiation source.
In their paper, the authors highlight the need for effective detection and monitoring of radioactive isotopes, citing events such as Japan’s 2011 Fukushima Daiichi nuclear power plant disaster and the threat of possible radiation releases from Ukraine’s war zone Zaporizhzhya. Machine learning-based algorithms and airborne radiation detection enable real-time monitoring of radiation incidents and integrated emergency planning, for example.
However, the authors also note that the daily operation of nuclear reactors, the mining and processing of uranium, and the disposal of spent nuclear fuel also require monitoring of radioisotope releases. In recent years, radiation localization has gained increasing interest in applications such as autonomous inspection of nuclear facilities.
In this study, the authors focused on gamma-ray sources, but the computational tools they developed to extract directional information from a limited number of pixels are not limited to specific wavelengths and can therefore be used for other forms of light, such as neutrons and ultraviolet light. The authors conclude that their framework provides a means to achieve high-quality radiation mapping with a simple detector configuration and is expected to be implemented in real-world radiation detection.
This article was first published in Nuclear Engineering International magazine.
