|
|
| Fig. 1: Pierre and Marie Curie in the laboratory, demonstrating the experimental apparatus used to detect the ionsation of air, and hence the radioactivity. (Source: Wikimedia Commons) |
Machine Learning (ML), a subfield of Artificial Intelligence (AI), refers to algorithms that are programmed to learn from observations and can then make statistical inferences based on what has been learned, rather than being based on the explicit knowledge of the programmer. [1] Deep Learning (DL), a subset of ML, defines algorithms that can analyse data with a logic structure similar to how a human brain would draw conclusions. In recent years, there has been great efforts in ML/DL development to enhance the detection of nuclear materials for regions where radioactive gauges, such as sodium iodine (NaI), may not perform well. This is most likely to occur when the material concentration is low and/or there are multiple isotopes mixed together in the sample of interest. [2] Newly developed ML and DL methods have shown promising performance in localising, quantifying, and identifying radioactive isotopes.
In the 1860s, R. Bunsen and G. R. Kirchhoff had developed a method of determining the chemical content of substances by heating them and analysing the spectra produced by the refraction of their flames; known as "spectroscopy". [3] The Curies, shown in Fig. 1, hoped that spectroscopy could provide some explanation to their revolutionary recordings of unexplained energies at their "ionising chamber" when testing "pitchblende". [3] Pitchblende had been valued principally as a source of Uranium (U). But when Marie Curie placed it in her condensation chamber, on February of 1898, she discovered that it produced current much more than that produced by Uranium (U) alone. [3] By this time, the Curies were operating on the hypothesis that pitchblende's dramatic ionising power was caused by an unknown element which is giving off energy much more than Uranium (U). [3] The problem with pitchblende, however, was that it contained an enormous variety of elements and could not be easily separated in the lab. The next step for the Curies was obvious, to isolate this hypothetical new element. [3] Despite several trials, the Curies failed to corroborate their findings through spectroscopy. They were unable to identify spectral lines for their mystery substance. [3] However, the fact was that they had obtained a substance that was 400 times as active as Uranium (U) with nothing comparable among known elements. They thus concluded that the substance extracted from the pitchblende contained a metal never known before. They named it Polonium (Po). [3] Fortunately, the Curies were able to attain spectral lines during their second phase of spectroscopy testing, which led them to the discovery of Radium (Ra). [3]
The same principles used by the Curies in identifying isotopes are still used today. Despite their unquestioned effectiveness, they are known to be costly and impractical when performing large scale isotope classifications. This has motivated the quest for automated methods to locate and identify isotopes.
ML/DL algorithms were thus brought to the problem, creating AI spectroscopy. AI studies in the nuclear chemistry field are scarce, yet they have significant potential. Radioactive hot-spot localisation and identification using DL was recently investigated by Mendes et al. [4] The work used artificial neural networks (NNs) to localise, quantify and identify radioactive sources. Both the radioactive source localisation and identification objectives were met. However there were limitations: [4] For localisation, accuracy is limited when the difference between source activities is high. For isotope identification, the requirement of large training data and the inability of identifying multiple isotopes are major drawbacks. [4]
|
|||||||||||||||
| Table 1: Samples containing pure isotopes that were mixed by Ayhan et al. to create test cases. [2] |
The performance of ML in detecting missing sources within a well-characterized sample was reported in Durbin et al. [5] A preliminary test sample, containing nine Cs-137 point sources arranged in a 3 × 3 grid, was simulated with the radiation transport code Monte Carlo N-Particle (MCNP6). [5] The simulations also included a detector array consisting of two NaI scintillators. [5] Simulations were performed with all sources present to characterize the sample. To simulate missing source scenarios, all combinations with one or two sources removed were also simulated. [5] The study proved that the k-nearest neighbour (KNN) algorithm can successfully predict the location of a missing source with 100% accuracy. Durbin et al. also claimed 99% accuracy when two sources were missing, but they did not provide corroborating evidence. [5] For the case of a single missing source, all 100 trials resulted in correct identification of the missing source. It is evident that the task in hand was not complicated for ML/DL incorporation, hence, the results were attained with high accuracy. It is included in this reporting for the sole reason of being one of the very few studies on AI in nuclear chemistry.
A more promising study on high performance remote radioactive material identification in mixtures of isotopes is depicted in Ayhan et al. and is of particular interest due its close resemblance to physical spectroscopy. [2] Several algorithms were studied to achieve enhanced detection of nuclear materials, starting with the traditional method of spectral analysis. The latter involves looking at certain regions of interest (ROIs) in the gamma (γ)-ray spectrum. [2] One drawback of the ROI approach is that it may not perform well when ROIs overlap significantly with the large libraries of radio-isotopes being used for reference. [2] ROI methods also do not use as much information as methods that utilize the entire spectrum. [2] Other methods to analyse spectral signatures from a mixture of materials are
|
||||||||
| Table 2: Average RMSE values of the test dataset (no background subtraction), from Ayhan et al. [2] | ||||||||
|
||||||||
| Table 3: Average RMSE values of the test dataset (with background subtraction), from Ayhan et al. [2] |
Partial Least Square (PLS), which has been used in rock composition analysis
Linear Regression (LR) and Random Forest Regression (RFR) are conventional ML tools used for un-mixing analysis. [2]
Mixtures of the samples shown in of Table 1 were made to emulate low Signal to Background Ration (SBR) conditions. [2] Simple linear mixing ws used. [2] Several thousand samples were generated to train the models. Since this may be considered low in the ML/DL community, it is a standard procedure to use cross-validation scheme to tackle this problem. [2] When applying Deep Regression (DR), fivefold cross validation training was applied. [2] After summing the synthetic mixture foreground and the appropriate background, Poisson noise was added to the spectrum. [2] The performance of the algorithms was measured using the concept of Root Mean Square Error (RMSE)
| RMSE | =[ | 1 N |
N ∑ i = 0 |
(Predicted MRi - True MR)2 | ]1/2 |
New results in using conventional ML and DL based algorithms for un-mixing multiple nuclear materials from mixed samples were attained through identifying radioactive isotopes in the mixtures. Three (PLS, RFR, and LR) out of six tested algorithms are highlighted. Experimental results demonstrated that PLS performed best in isotope identification under both low and high concentration of material, making it a candidate for algorithm generalisation (see Tables 2 and 3). It is also worth mentioning that the results are as reliable without background subtraction, which increases implementation practicality. Further improvements to PLS performance could be achieved by applying deconvolution algorithms for training, as this would allow direct estimation of peak locations and then use of the peak location information to identify components in the mixture. Implementation of the the latter is not published to date.
AI applications in nuclear chemistry are young and few are comprehensive enough for reliable application. Reasoning and results displayed in Ayhan et al., are particularly promising. [2] RMSE values proved that most selected algorithms function well in identifying the comprising individual isotopes mixed linearly in an actual sample. This could be revolutionary in gearing actual spectroscopy to AI spectroscopy in the field of nuclear chemistry.
© Nora AlMaqsseed. The author warrants that the work is the author's own and that Stanford University provided no input other than typesetting and referencing guidelines. The author grants permission to copy, distribute and display this work in unaltered form, with attribution to the author, for noncommercial purposes only. All other rights, including commercial rights, are reserved to the author.
[1] C. F. Uribe et al., "Machine Learning in Nuclear Medicine: Part 1 - Introduction," J. Nucl. Med. 60, 451 (2019).
[2] B. Ayhan et al., "High-Performance Remote Radioactive Material Identification of Mixtures," IEEE 9625926, IEEE Trans. Nucl. Sci. 69, 86 (2022).
[3] S. Quinn, Marie Curie A Life (Simon and Schuster, 1995).
[4] F. Mendes et al., "Radioactive Hot-Spot Localisation and Identification Using Deep Learning," J. Radiol. Prot. 42, 011516 (2022).
[5] M. Durbin, A. Kuntz, and A. Lintereur, "Machine Learning Applications for the Detection of Missing Radioactive Sources," 2019 IEEE Nuclear Science Symposium and Medical Imaging Converence (NSS/MIC), IEEE 9059881, 265 Oct 19.
