Automation of etch pit analyses on solid-state nuclear track detectors with machine learning for laser-driven ion acceleration.

Solid-state nuclear track detectors (SSNTDs) are often used as ion detectors in laser-driven ion acceleration experiments and are considered to be the most reliable ion diagnostics since they are sensitive only to ions and measure ions one by one. However, ion pit analyses require tremendous time and effort in chemical etching, microscope scanning, and ion pit identification by eyes. From a laser-driven ion acceleration experiment, there are typically millions of microscopic images, and it is practically impossible to analyze all of them by hand. This research aims to improve the efficiency and automation of SSNTD analyses for laser-driven ion acceleration. We use two sets of data obtained from calibration experiments with a conventional accelerator where ions with known nuclides and energies are generated and from actual laser experiments using SSNTDs. After chemical etching and scanning the SSNTDs with an optical microscope, we use machine learning to distinguish the ion etch pits from noises. From the results of the calibration experiment, we confirm highly accurate etch-pit detection with machine learning. We are also able to detect etch pits with machine learning from the laser-driven ion acceleration experiment, which is much noisier than calibration experiments. By using machine learning, we successfully identify ion etch pits ∼105 from more than 10 000 microscopic images with a precision of ≳95%. A million microscopic images can be examined with a recent entry-level computer within a day with high precision. Machine learning tremendously reduces the time consumption on ion etch pit analyses detected on SSNTDs.

3D track extraction from a fluorescent nuclear track detector via machine learning and an application to diagnostics of laser-driven ions.

We have developed an ion diagnostic method for laser-driven ion acceleration experiments that uses fluorescent nuclear track detectors (FNTDs). An FNTD records the particle tracks as color centers and does not require chemical etching, unlike CR-39 track detectors. The color centers are observed using a confocal laser microscope, and 3D particle tracks can be obtained by changing its focal position. The intensity of the color centers corresponds to the energy deposited by the ions. The nuclides of the ions can be determined from the intensity distribution of the color centers as a function of depth and the distance between the stopping point and the surface of the detector. To extract the intensity distribution, we must track the same ion tracks in the depth-layered microscopic images from the surface to the stopping point, even if they overlap with those of other ions. In addition, since an FNTD is sensitive not only to ions but also to electrons and photons, we must identify ion tracks among those from the latter particles. To analyze a statistical number of ion tracks, it is necessary to automate these processes. We have thus developed a method for automated ion detection and 3D tracking that relies on a support vector classifier and a kernelized correlation filter. This method was tested on a laser ion acceleration experiment performed using the J-KAREN-P laser. The method automatically detects ion tracks on FNTDs and tracks them in the depth direction. The training data are sampled from the Heavy-Ion Medical Accelerator in Chiba.