Quantifying Uncertainty in Source Term Estimation with Tensorflow Probability

Abstract

Fast and accurate location and quantification of a dangerous chemical, biological or radiological release plays a significant role in evaluating emergency situations and their consequences. Thanks to the advent of Deep Learning frameworks (e.g. Tensorflow) and new specialized hardware (e.g. Tensor Cores), the excellent fitting ability of Artificial Neural Networks (ANN) has been used by several researchers to model atmospheric dispersion. Despite the high accuracy and fast prediction, regular ANNs do not provide any information about the uncertainty of the prediction. Such uncertainty can be the result of a combination of measurement noise and model architecture. In an urgent decision making situation, the ability to provide fast prediction along with a quantification of the uncertainty is of paramount importance. In this work, a Probabilistic Deep Learning model for source term estimation is presented, using the Tensorflow Probability framework.