Researchers from the Institute of Computer Science at the University of Tartu (Estonia), together with the technology company GScan, are developing artificial intelligence (AI)-based methods that help more accurately determine the location of objects from data collected using cosmic particles. The technology could be useful for assessing the condition of bridges and other large structures and identifying possible damage.
This concerns muon tomography, a scanning method that uses muons, natural and harmless particles produced when cosmic rays collide with the atmosphere. They can pass through meters of concrete, steel and rock. Detectors record each muon on entry to and exit from an object. Most particles travel straight, but those that encounter something dense along the way are slightly deflected. These tiny deflections make it possible to see the internal structure without opening it up and without the artificial radiation required for an X-ray scanner.
The difficulty is that muons arrive slowly and the measurements are noisy, so images often come out blurry. This is precisely the problem the researchers are addressing with AI: they teach computers to build a clear picture from imperfect data. "Processing muon data requires complex solutions in machine learning and data science," says GScan product development engineer Tarvo Metspalu. "Research at the university helps us avoid dead ends in product development and improve the accuracy of the technology so that we can assess the internal condition of buildings and structures without damaging them."
Lecturer Kallol Roy and his colleagues noticed that muons that pass through an object without deflection carry no less information than deflected ones. A straight path indicates that there is nothing dense along it, while a deflected path indicates that something is there. Some data show where an object is not, while other data show where it most likely is. The researchers built machine learning models that read both types of clues simultaneously. In computer simulations, these models determined the location of hidden objects more accurately than models using only one type of data, and more accurately than the standard method used in muon imaging. The results were published in the journal IEEE Access.
In addition, the scientists developed methods for distinguishing closely located objects that may look like one large object in muon images. Their hybrid method reduced the proportion of erroneously merged objects in simulations from about 15% to 2%. "This result is important because in practical applications we need to be able to distinguish the location of objects and finer details, as well as detect potential changes inside structures," Roy added.
Besides determining location, the researchers are studying the possibility of identifying different materials from muon measurements. So far they have analyzed the properties of 14 materials and tested statistical methods and machine learning methods to learn to distinguish them. If AI can determine not only the location of an object but also its material, the technology could prove useful for inspecting industrial facilities or detecting hazardous and prohibited substances.
The next goal is to improve the reliability of the models and refine the datasets on which they are trained so that AI can work with a wider range of materials and more complex measurement conditions. Although further development and testing are still required for the technology to be widely used in assessing the condition of buildings and structures, the collaboration between the University of Tartu and GScan shows how combining physical measurements and AI opens new possibilities for inspecting structures whose internal makeup was previously difficult to verify.