Determinación del espectro de neutrones mediante redes neuronales artificiales en CPU y GPU
The neutron spectrum extends at several energies, so the counter used is the Bonner Spheres Spectrometer (BSS), using counting rates and Artificial Neural Networks (ANNs), prove to be an alternative method in neutron spectrometry. The CPU is limited to computationally intensive calculations. So a Gr...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Estado: | Versión publicada |
| Fecha de publicación: | 2017 |
| País: | México |
| Institución: | Universidad Autónoma de Zacatecas |
| Repositorio: | Repositorio Institucional Caxcán |
| Idioma: | español |
| OAI Identifier: | oai:http://ricaxcan.uaz.edu.mx:20.500.11845/1349 |
| Acceso en línea: | http://ricaxcan.uaz.edu.mx/jspui/handle/20.500.11845/1349 |
| Access Level: | acceso abierto |
| Palabra clave: | CIENCIAS FISICO MATEMATICAS Y CIENCIAS DE LA TIERRA [1] espectro de neutrones esferas Bonner redes neuronales artificiales GPU MATLAB neutron spectrum Bonner spheres artificial neural networks |
| Sumario: | The neutron spectrum extends at several energies, so the counter used is the Bonner Spheres Spectrometer (BSS), using counting rates and Artificial Neural Networks (ANNs), prove to be an alternative method in neutron spectrometry. The CPU is limited to computationally intensive calculations. So a Graphics Processing Unit (GPU) is attractive for computing with ANN, since it works in parallel. This study determined the neutron spectrum from the 7 counting rates obtained from the BSS using an ANN-trained CPU and NVIDIA® GPU. Neutron spectra were obtained from the International Atomic Energy Agency (IAEA) database. The counting rates of the BSS and the spectrum are related through the Fredholm equation which is a poorly conditioned system. To solve the problem an ANN feedforward was designed, consisting of 7 inputs, 2 hidden layers and an output of 25, 25 and 27 neurons. For the network training 182 spectra were taken, the values of the synaptic and bias weights were updated using the gradient conjugate descending algorithm (SCG). For the validation the remaining 12 spectra were taken and the spectra reconstructed by the ANN with the originals were compared using the Chi Square test 2 . The design was done with the neural network and parallel computing toolbox, MATLAB® 2015a. The training was performed in CPU with one and several cores, in CPU with GPU, and in GPU. The computational performance of the ANNs is better with the SCG algorithm, but on the contrary, it requires more memory capacity. The bottleneck in processing between CPU and GPU is the transmission speed in the PCI-E duct. |
|---|