Development of a Deep Rectifier Neural Network for dose prediction in nuclear emergencies with radioactive material releases

A Recent investigation showed that a 5-layer multilayer perceptron (5L-MLP) artificial neural network (ANN) accurately predicted the spatial dose distribution for accident scenarios with radioactive material release in Nuclear Power Plants (NPPs). However, the training took 04:30:00 h for a simplified scenario. This imposes limitations on the 5-MLP to deal with more realistic situations. In order to overcome such limitation, this work proposes the use of Deep Rectifier Neural Networks (DRNNs), running on a GPU-based parallel framework. The DRNNs were evaluated considering two realistic accident scenarios simulated with the atmospheric dispersion system used in a Brazilian NPP. In the more simplified scenario, the best DRNN achieved an error 25% lower than the 5-MLP with a training 155 times faster. In the more complex scenario, the best DRNN achieved an average error of 0.0213 with a training time of 30 min, demonstrating that DRNNs can improve ANN-based dose prediction in realistic situations. •Deep learning model able to accurately predict radiological plume dispersion.•The model is able to predict dose dispersion up to 1 h after complex accidents.•The model is able to improve the performance of mobile dose prediction systems.