Quantitative alpha spectra isotope identification based on deep learning method and Monte Carlo modeling

It is essential to identify radioactive isotopes contained in samples quantitively for environment radiation contamination monitoring. Energy-spectra-based nuclide identification is the most commonly used method in environmental radiation measurement and has been widely used for gamma ray emitting isotopes. In this work, a complete procedure of quantitative alpha spectra analysis and alpha isotope identification based on deep learning method was proposed and it was proved effective and feasible to use Monte Carlo simulation in deep learning model development. Deep learning model with 6 hidden layers (about 0.7million trainable parameters) was developed and implemented based on Keras framework in Python. Instead of complicated spectrum feature extraction algorithm, proposed deep learning model used spectrum data as input directly and could output a weight vector of nuclides with elements - -having - meaning of the percentage content of every nuclide. To train the proposed model, huge number of training spectra were needed and it was not possible to prepare all needed data by measurement. In this study, Monte Carlo simulation was used as an alternative method to produce massive alpha spectra by mixing basic spectrumof single nuclide. Based on the setups of commercially available Canberra alpha meter, a set of alpha spectra from single nuclide were generated by Geant4 simulation. To broad representativeness of simulated spectra, different measuring conditions were taken into consideration and more than 500 basic spectra were prepared to form a data library. Randomly choosing basic spectra from the library were summed with random-weight to generate millions of training spectra. The proposed model was trained using 4 million generated spectra - and tested by 3 kinds of spectra, namely spectra from generation-based spectra library, Geant4 simulation of mixed nuclides and measurements of actual mixed alpha source. The identified percentage contents by model prediction were compared with the original weight values to validate qualitative and quantitative identification of nuclides contained in spectra. Absolute errors of identified nuclide percentage contents were less than 1% for simulated spectra and less than 2% for acutal measured spectra. The accuracy could be improved further by increasing the number of nuclides contained in library, the range of measuring conditions in simulation and the data set of generated training spectra used in model training. Results indicated that deep learning mode could identify the nuclide contained in complicated alpha spectra and determine the content of nuclide with good accuracy. The method proposed in this study has great potential in quantitative spectral alpha isotope identification.