Machine-learning could help us use cosmic muons to peer inside large objects such as nuclear reactors. Developed by researchers in China, the technique is capable of identifying target materials such as uranium even if they are coated with other materials.
The muon is a subatomic particle that is essentially a heavier version of the electron. Huge numbers of cosmic muons are created in Earth’s atmosphere when cosmic rays collide with gas molecules. Thousands of cosmic muons per second rain down on every square metre of Earth’s surface and these particles can penetrate tens to hundreds of metres through solid materials.
As a result, cosmic muons are used to peer inside large objects such as nuclear reactors, volcanoes and ancient pyramids. This involves placing detectors next to an object and detecting muons that have passed through or scattered within the object. Detector data are then processed using a tomography algorithm to create a 3D image of the object’s interior.
Illicit nuclear materials
Muons tend to scatter more from high-atomic-number materials, so the technique is particularly sensitive to the presence of materials such as uranium. As a result, it has been used to create systems for the detection of illicit nuclear materials hidden in freight containers.
Muon tomography is relatively straightforward when the object is of simple construction – such as a pyramid built of stone and containing voids. Producing useful images of more complex target – such as a freight container full of unknown objects – is much more difficult. The conventional computational approach is to calculate the muon-scattering physics of many different materials and combine these data with muon-tracking algorithms. This, however, tends to require huge computational resources.
Supervised machine learning has been used to reduce the computational overhead, but this requires prior knowledge of the target materials – limiting efficacy when imaging unknown and concealed materials. What is more, many materials in complex objects are coated with other materials and these coatings can affect muon scattering.
Now, Liangwen Chen at the Institute of Modern Physics of the Chinese Academy of Sciences and colleagues have used a technique called transfer learning to improve cosmic muon tomography of objects that contain coated materials. The idea of transfer learning is to begin with knowledge of the muon-scattering parameters of bare, uncoated materials and use machine learning to predict the parameters of coated materials. Chen and colleagues believe that this is the first application of transfer learning to muon tomography.
Monte Carlo simulations
The team began by creating a database describing how cosmic muons interact with representative materials with a wide range of atomic numbers. This was done by using Geant4 to do Monte Carlo simulations of how muons interact as they pass through materials. Geant4 is the most recent incarnation of the GEANT series of computer simulations, which have been used for over 50 years to design particle detectors and interpret the data that they produce.
Chen and colleagues used Geant4 to calculate how muons are scattered within nine materials ranging from magnesium (atomic number 12) to uranium (atomic number 92). These included common elements such as aluminium, copper and iron. The geometry of the scattering involves incoming cosmic muons with energies of 1 GeV and incident angles that are typical of cosmic muons. After scattering from a material target, the simulation assumes that the muons travel though two successive detectors, which measures the scattering angles. Data were generated for bare targets of the nine materials, as well as the nine materials coated with aluminium and polyethylene. Each simulation involved 500,000 muons passing through a target.
These data were then sampled using an inverse cumulative distribution function, as well as integration and interpolation. This is done to convert the data to a form that is optimal for training a neural network.
Muons: probing the depths of nuclear waste
To use these data, the team created two lightweight neural-network frameworks for transfer learning: one based on fine tuning; and the other a domain-adversarial neural network. According to the team, both frameworks were able to identify correlations between muon scattering-angle distributions and different target materials. Crucially, this was the case even when the target materials were coated in aluminium or polyethylene.
Chen explains, “Transfer learning allows us to preserve the fundamental physical characteristics of muon scattering while efficiently adapting to unknown environments under shielding”.
Chen and colleagues are now trying to apply their process to more complicated scattering geometries. The also plan to include detector effects and targets made of several materials.
“By integrating simulation, physics, and data-driven learning, this research opens new pathways for applying artificial intelligence to nuclear science and security technologies,” says Chen.
The research is described in Nuclear Science and Techniques.
