Conference Agenda

Precise and reliable particle detection is an essential aspect of many fields, such as energetics, healthcare, industry, science, and research. Research in particle detection focuses on the development of methods that allow for the precise particle recognition, together with minimizing the uncertainty of the measurement. The presented work focuses on the processing of the signal from the Solid-State Nuclear Track Detectors (SSNTDs), which are widely used in ion and fast neutron detection. The particles disrupt the material of the detector, creating latent tracks, which are subsequently enlarged by etching. The etched detectors are usually analyzed using the optical microscope, where the associated dosimetric quantity is directly proportional to the number of etched tracks per unit of area. Except for the etched tracks, the surface of the SSNTD can also contain bubbles, dust, or microscopic scratches, representing the background signal.
In the past, the SSNTDs were analyzed manually; however, technological development allows for the incorporation of the automation of the scanning and analysis. An example of such an automated system is the TASLImage system (Tasl) used at the Slovak Technical University in Bratislava (STU), which was developed to analyze the plastic SSNTDs of TASTRAK type. The system enables the evaluation of the measurements of fast neutrons and the measurements of the radon concentration, and provides the user with the result values, photos of the detectors, and the parameters of the registered events.

Artificial intelligence methods have been gradually applied in the field of SSNTDs, particularly through the use of object detection networks. These models enable the detection of objects with high accuracy. However, detecting small objects remains a separate challenge, as performance tends to be significantly lower compared to larger objects. Despite this, ongoing advancements are steadily improving the accuracy of small object detection as well. The task addressed in this work falls within this category, where YOLO-based detectors have demonstrated particularly strong performance.

In this work, we propose an automated approach for particle track detection using deep learning-based object detectors, specifically the YOLO (You Only Look Once) family models, trained to identify and localize etched tracks in detector images. YOLO's real-time object detection capabilities, combined with its proven performance in scenarios with high object density and variability, make it an attractive candidate for this application. We aim to explore the potential of YOLO-based analysis as a faster and more flexible alternative to Tasl's performance. This approach is motivated by the need for scalable, field-deployable, and explainable detection systems suitable for laboratory and on-site applications.

The YOLO-based analysis was performed on the dataset created at STU, which includes the images of the surfaces of the etched TASTRAK detectors from various experiments, annotated with bounding boxes corresponding to visible etched tracks identified based on Tasl output. These images capture various track morphologies, reflecting different particle energies and incident angles. The inaccuracies in detecting very small objects around 10 pixels were observed in the Tasl analysis, together with the rejection of partial areas of the images. The images that contained areas omitted during the analysis performed by Tasl were excluded from the training dataset and a filter was applied to exclude events smaller than 10 pixels in area (corresponding to 5 μm).
The final dataset consists of 18,692 images, which we split into training, validation, and test sets with 14,953, 1,870, and 1,869 images, respectively (following an 80:10:10 ratio). As annotations for etched tracks, we use bounding boxes based on the Tasl decision.

In our experiments, we have trained YOLOv5 and YOLOv11 models and evaluated their performance on unseen images (test split) using standard object detection metrics such as precision, recall, and mean average precision (mAP) at different intersection-over-union (IoU) thresholds. The standard mAP for small object detection is around 20-50%, depending on visibility and contrast between the background and the object. To benchmark the effectiveness of our method, we compare YOLO detections against Tasl identifications, analyzing the consistency of track counts, spatial localization accuracy, and false positive/negative rates. We also assess the time efficiency of detection to highlight YOLO’s potential advantage in processing speed.

Preliminary testing indicates that the YOLO detectors can identify a range of etched particle tracks and can be compared to the Tasl. The final paper will present a detailed detection performance evaluation, including sensitivity, precision, and processing efficiency. Filtering the dataset improved the mAP@0.5 from 52% to 60%. These results suggest that further improvements are possible with additional correction of the dataset annotations.

This research aims to demonstrate the feasibility of deep learning-based approaches for enhancing the efficiency, scalability, and accessibility of particle track detection in radiation and particle physics applications. Ongoing and future work will focus on increasing model robustness through data augmentation and exploring instance segmentation methods for finer morphological analysis of particle tracks.

The computational cost of nuclear safety thermal-hydraulic analysis can become a burden if there is a large number of calculations or a quick need for results. In this aspect metamodels appear as an enough accurate approach that can reduce computational cost.

This project presents metamodels based on Neural networks architectures to preform faster predictions of relevant variables in the context of the Large Break Loss of Coolant Accident for a Pressurized Water Reactor. These techniques have been developed on the field of the Machine Learning and Deep Learning, which have been getting more relevance on last decades addressing the long-term dependencies issue.

The models, trained on data from TRACE simulations, are able to predict variables, such as cladding temperature, local oxidation or liquid level, predicting its entire time series or just one value among all the time series timesteps. The entire series prediction provides the possibility of complete by-passing the thermal-hydraulic code for analysing just these variables, if the available time is minimal.

The neural network metamodel performs the predictions of the full peak cladding temperature behaviour with errors below 25 K during the LBLOCA.

Simulations play a vital role in understanding complex phenomena during the late stages of reactor accidents, such as the formation and cooling of debris beds. Cooling these debris beds, which may form either inside or outside the reactor pressure vessel (RPV), is critical for mitigating severe accident consequences. The phenomena that occur during the progression of debris bed cooling are highly complex and are computationally expensive to model realistically [1]. The application of Artificial Intelligence (AI) has proven invaluable in accelerating simulations and identifying correlations between variables, offering new insights and reduced computational time compared to traditional methods. Hence, as one of five doctoral research projects in the InnoPhase project, funded by the German Federal Ministry of Education and Research (BMBF), an application of machine learning (ML) methods is being investigated for the surrogate modelling of computationally expensive debris bed cooling simulations.

This research focuses on applying AI to COCOMO (Corium Coolability Model) simulation data, validated on experimental data from the FLOAT facility [2], developed at the Institute of Nuclear Technology and Energy Systems (IKE), University of Stuttgart [3]. COCOMO simulates two-phase flows of steam, non-condensable gases, and water to predict debris bed coolability under varying thermal-hydraulic conditions and is used in this research to simulate the cooling of a hot debris bed in water to saturation temperature – also known as quenching. Initial AI methods were applied to pre-existing 2D simulations of the FLOAT facility, where artificial neural networks (ANNs) demonstrated the ability to reduce the prediction time for quenching from minutes to milliseconds [4]. This work also served as a proof of concept for applying AI to more complex debris bed geometries in future studies.

Furthermore, comprehensive 2D ex-vessel debris bed database has been developed, consisting of 5,572 data points across 9 uniformly distributed input parameters, including corium composition, particle size, porosity, system pressure, initial bed temperature, and bed geometry. The database contains approximately 4633 quenched and 939 melted scenarios. Various AI techniques were employed and compared to predict the simulation outcome of whether the debris bed quenched or melted. Among these methods, the 'Voting Classifier' achieved an accuracy of 99.3%. This approach was further extended to predict the total quench time and the location of the quench front within the debris bed. The ANN model demonstrated an average Structural Similarity Index (SSIM) of 0.8831, reflecting strong similarity with the simulation data, while also significantly reducing computational time. In comparison to COCOMO simulations, which can take up to 10 hours per scenario, the pre-trained NN produces predictions in milliseconds, with training requiring up to 30 seconds. This reduction in computation time highlights the NN's potential for real-time applications in safety assessments.

Additionally, Active Learning (AL) methods developed by another participant of the InnoPhase project [5] are being implemented with COCOMO to significantly reduce the amount of necessary data for a robust and accurate model.

This project is funded by the Federal Ministry of Education and Research (BMBF) under the funding code 02NUK078B. The training data was determined using the IKE detail code COCOMO.