Radionuclide classification based on gamma spectroscopy and artificial intelligence

Currently, in almost all segments of the production chain, automation is a requirement for productivity improvement. With respect to nuclear facilities, active online monitoring is one of best practices for nuclear security and safety maintenance, to prevent incidents that could compromise a particular installation. In this context, spectral signature monitoring automation can be explored, aiming at the rapid identification of adverse events, such as radiological accidents. The main objective of this work was an automated radionuclides classification technique establishment, using an Artificial Neural Networks (ANN) architecture. The methodology used consisted basically of simulating the geometry of an established experimental apparatus, using the MCNP5 code, obtaining the simulated gamma spectral signature for the studied nuclides. The simulated spectra were used to compose the ANN training and testing data set, while the experimental spectra were subjected to the artificial intelligence model classification, in order to allow the neural network quality assessment. The final developed architecture of ANN was correct to recognize the experimental spectra of 60Co, 137Cs and 152Eu. Therefore, the results were satisfactory and proved automation technique development viable.


INTRODUCTION
Currently, there are constant changes in labor relations and in the productivity increase in the industry and services sectors, based on artificial intelligence solutions. Regarding the nuclear installations area, there are points that deserve attention, especially those dealing with these installations safety, as critical events such as leaks, radioactive elements dispersion in the atmosphere and even nuclear material theft may, not only compromise physical safety of a plant, but also exposing the public to hazardous conditions. In this context, artificial intelligence, especially supervised machine learning, can offer an agile way of detecting security incidents, allowing for quick action by the team responsible for the installation.
The main purpose of this work is to establish an automated classification model for radionuclides based on Artificial Neural Networks (ANN). The basic study concept is to simulate the gamma spectral signature of the nuclides studied set, through the MCNP5 code, to compose the ANN training and tests set. Then, the model is validated with data obtained from the experimental survey of the spectra of some of these nuclides.
Studies with models similar to the one proposed, using computational simulation based on the Monte Carlo method for ANN training, which aim at spectra classification, appear in the literature already in the nineties, as in the work of FUKUDA and KITAMURA [1] . In this study, the realtime analysis of spectra was sought for nuclear safety purposes. In fact, a functional model was arrived at, but it was concluded that there was still a need for additional studies due to the incipience of the technique at the time.
In 2005, the work of NUNES [2] used the ANN training model based on simulation with the use of the MCNP, for detection of landmines, based on the recognition of characteristic gamma spectra patterns, resulting from the activation of the materials present in the explosives by thermal neutrons.
The result obtained shows that not only the ANN was able to recognize the patterns of the soil compositions where it was trained, but also acted in the identification of the presence of mines in situations outside this scope.
In 2015, the work of VARLEY [3] used an ANN application, with training based on simulation data with the MCNP, in order to identify traces of 226 Ra, resulting from military activities in the United Kingdom. It is concluded that the technique provided a reliable means for estimating the activity of point sources and information on possible layers of soil contamination.
Recently, in 2020, JHUNG et al. [4] used an ANN to identify gamma-emitting radionuclides using silicon photomultipliers. In this case, the training was carried out with experimental spectra, different from the spectra submitted to classification. It was concluded that the network was able to classify them correctly, and that increasing the depth of the network can improve the classification results.
In this context, and in addition to the main proposal, the study led to the possibility of enabling a model that could support an alternative project to the use of detectors available on the market, which despite offering the functionality of spectra classification, are proprietary technology. Thus, the artificial intelligence model used in this work will be able to establish the pillars of a software that can compose the core of an industrial automation system, which, in turn, aims at the automated monitoring of the spectral signature of a nuclear installation.

MATERIALS AND METHODS
In order to facilitate the understanding of the various steps of the experiments carried out, this study was divided into two phases. The first phase consisted of defining the experimental apparatus, reproducing this apparatus in the computational environment of the MCNP5 [5] code, comparing the simulated and experimental spectra of 137 Cs, 60 Co and 152 Eu, and validating the initial configuration of the ANN. The second phase consisted of expanding the nuclides present amount in the training and testing base, and adjusting the ANN parameterization, so that it produces results consistent with this new knowledge base.
The Neural Network implementation was made using the Python programming language [6], Keras [7] and Tensorflow [8] frameworks, used within Google's collaborative computing environment, Colab [9]. The main metric for monitoring the performance of the ANN versions used in the experiments performed was the loss function, using the cross entropy algorithm [10], with the accuracy function as a secondary metric.

First phase
The equipment used to assemble the experimental apparatus consisted of: (i) A sealed source of  [11]) and on the work of MEDEIROS [12]. The uncertainty associated with geometric measurements is 0.1 mm. The experimental spectra obtained were used in the calibration of the detection system, with the variability of the uncertainty associated with the count statistics being within acceptable limits (given by n 1/2 / n, where n is the counts number of a given spectrum, in the region that forms the Gaussian of each of its photopeaks). After the experimental survey of the 137 Cs, 60 Co and 152 Eu spectra, the experimental apparatus was reproduced within the computational environment of MCNP5.
The simulation was performed on a computer with an Intel(R) Core(TM) i7-8700 3.2 GHz processor (6 physical cores and 12 logical cores), with 8 Gb RAM memory installed, running Windows 10 Home operating system 64-bit. In the simulations, 385 million stories were used for each nuclide, with an average simulation time of 10 hours, using 10 cores.  From the simulation model, the simulated spectra of the aforementioned nuclides were obtained, using the Tally F8 and the GEB function (Gaussian Energy Broadening, a function that simulates the energy resolution of a real radiation detector [5], using the parameters a, b and c with the estimated values at -9.3408 x 10 -3 , 7.5107 x 10 -2 and 0.5843 respectively). Once the experimental and simulated spectra were obtained, a visual analytical comparison was performed, which consisted of the graphic overlay between these spectra.
After the graphic analysis described, the spectra obtained from the simulation model were used   Co, 54 Mn), whose spectra were obtained from simulation with the MCNP5. This inventory was selected based on the studies by CURZIO [14], who developed a radioactive plume dispersion analysis resulting from a simulated accident with a small PWR reactor and PEREIRA [15], which analyzes a hypothetical dirty bomb explosion at a large public event. The radionuclides stored in the laboratory of the Instituto Militar de Engenharia (IME) were kept in the inventory, namely 152 Eu, 133 Ba, 57 Co, 60 Co and 54 Mn, in addition to 137 Cs, present in the inventory surveyed by CURZIO [14].
The experimentally obtained spectra of 137 Cs, 60 Co and 152 Eu were again used as ANN production data.
In this phase, the ANN configuration was started with the same final configuration of the first phase, which, when submitted to a new set of training data and tests, did not obtain satisfactory performance. A systematic evolution was carried out with alteration of several ANN parameters, having as indicators of improvement the curves of the loss and accuracy functions, as mentioned above. The final configuration obtained in this phase was an ANN with four layers, obtained by inserting a hidden layer before the output layer of the previous configuration, with 32 neurons and using the sigmoid activation function. In addition, hyper parameters such as learning rate, learning batch size [16], and the optimization algorithm were changed, the latter causing a significant improvement in the results, using ADAM [17] from this phase on.

RESULTS AND DISCUSSION
The work had as its starting point the ANN architecture of the study by NUNES [2], differing in the application of the classifier, obtaining assertiveness similar to the work cited. Thus, a neural network capable of distinguishing between the spectra submitted to classification within the set of radionuclide spectra that composed the training base of the network was obtained.
The comparison with the studies of VARLEY [3] and JHUNG et al. [4] shows that, like the aforementioned authors, this study shows the feasibility of using Artificial Neural Networks for the classifiers implementation within the nuclear area.
Next, the results obtained will be described, as well as the relevant discussions, referring to each experiment phase in the respective subsequent sections.

First phase
Once the experimental apparatus was defined and assembled, the energy calibration of the detection system was performed [18,19] with the nuclides 60 Co, 152 Eu and 137 Cs. Figure 3 shows the visual comparison between the graphically superimposed experimental and simulated spectra for the defined experimental apparatus.
A visual adhesion, illustrated in Figure 3, was observed between the experimental and simulated spectra for 3 used nuclides. However, there are some deviations between the overlapping spectra, which may require a detailed analysis in a future study.
Once the simulation results were validated with the data obtained experimentally, the Neural Network used as the main point of the automated radionuclide classification system was parameterized. Thus, after using the analytical comparison to validate the computational simulation model, the simulation data, as well as the experimental spectra, were submitted to classification with the ANN configured in the first phase of the experiment. The loss and accuracy curves obtained from this classification are illustrated in Figure 4. The classification result is shown in table 2.

Second phase
The results obtained in the second phase of the experiment, with the four-layer ANN configured as described in section 2.2, relating to the curves of the loss and accuracy functions are illustrated in   It was possible to observe that the curves of the loss and accuracy functions quickly tended to zero and one respectively, over the training epochs, with the final loss being 1.5170 x 10-4 in the first phase and 2.335 x 10 -5 with respect to the second phase. In addition, all pertinence relations, obtained through the classification of nuclides used as production data, returned results above 0.98, which would perfectly suit a practical application of the study.

CONCLUSION
The study carried out had as its main proposal to establish an automated classification model for radionuclides, based on the Artificial Neural Networks application and supervised machine learning. In fact, it was described how the application of ANN can produce viable results for the development of an automation technique, in order to meet this project requirement.
Furthermore, having a path that can be developed in order to obtain a viable alternative to detectors with similar functionality, high cost and available only from international manufacturers, contributes to the strengthening of national industry and technology with regard to the applications of radiological defense.
It should also be noted that the implementation of the ANNs in this study was carried out with free and open source technologies, such as the Python language, and the Keras and TensorFlow libraries, which brings the versatility needed to make the developed models part of a customized automation system .
A continuation of this study may aim at an automation system with a core based on artificial intelligence, which can be configured, for example, to monitor the spectral signature of a specific type of nuclear installation, aiming at increasing its radiological safety.
Another future study, derived from this work, would be the use of Deep Learning, as suggested in the study by JHUNG et al. [4], with the substantial increase of layers in the used neural network, which could serve as a work proposal whose objective would be to compare the results to be obtained with the results of the present study.