Real-time radiation measurements using neural networks

The observation of light is a cornerstone of astronomy, as the light emitted by celestial objects carries vital information about their composition, temperature, and motion. Light spans a vast range of wavelengths, from radio waves to gamma rays, with each range revealing unique insights into the universe. However, a portion of the electromagnetic spectrum, the medium-energy X-rays and gamma rays known as the “MeV gap”, remains challenging to observe. Existing detection technologies struggle to provide accurate and efficient measurements in this range, leaving a significant blind spot in our understanding of high-energy cosmic phenomena.

A new radiation sensor based on cadmium zinc telluride (CZT) semiconductor material has been designed at DTU Space. This sensor can detect radiation within a three-dimensional volume with sub-millimeter precision. The innovative technology has the potential to revolutionize multiple fields of science and industry. However, a significant challenge remains in enhancing the sensor’s capability to rapidly and accurately determine the exact location and energy of radiation sources. Traditional computational methods often struggle to efficiently process the high-dimensional, complex data generated by these sensors.

My PhD research focuses on integrating machine learning and deep learning computational methods with existing CZT sensor technology to enhance its predictive capabilities. By leveraging data-driven models tailored to the specific requirements of radiation detection, this work aims to significantly improve the sensor’s performance and usability. Key objectives include achieving greater accuracy in locating radiation sources and reducing the computational time required for analysis. Ultimately, this research seeks to advance the technological readiness level of CZT sensors, enabling their deployment in demanding applications such as space exploration.

The primary challenges of this research project arise from the complexity of the data and the lack of established ground truth for training machine learning and deep learning models. To address these hurdles, the core approach integrates known physical principles governing the behavior of electrons within the sensor with advanced data-driven techniques. By combining physics-based models with machine learning, the goal is to develop a hybrid framework capable of extracting meaningful patterns from complex sensor data.

This approach not only addresses the current limitations of radiation detection but also paves the way for innovations in applications such as space exploration, medical imaging, and beyond. By merging physics with data science, this research aims to bridge gaps in understanding while driving the development of smarter and more reliable technologies for the future.