Vented gas explosion overpressure calculation based on a multi-layered neural network

The case of a gas explosion occurring in a geometrically simple enclosure, equipped with a vent is considered. It is well known in the gas explosion scientific community that the calculation of the reduced explosion overpressure, determinant in safety studies, is not trivial. Not only there is a strong dependency on the chemical kinetics of the combustible but also on the enclosure geometry, the fluid flow, the vent mechanical behaviour, shape, etc … As a result, the modelling of the physics at stake is challenging, a wide range of models are proposed in the scientific literature and this reference situation is still the object of extensive research. A new simulation approach ignoring a large part of the underlying physics is investigated. It is based on the use of an artificial neural network (ANN). The focus is given on the method of use and results obtained with the ANN rather than on the neural network itself. Our observations are discussed within the scope of industrial safety problems. Calculations performed with the relatively simple ANN proposed in the official TensorFlow tutorial, on a vented explosion database containing 268 tests, led to surprisingly good results considering the ANN implementation efforts. The tool might look promising but is also far from being as trivial as it seems at a first glance: not only the results of simulations obtained with this type of model must be examined with the greatest care but also the initial data base must be very well controlled. Routes are proposed to enhance the initial database and perform relevant analyses of the neural network predictions. •Vented explosion overpressures are examined with an artificial neural network (ANN).•The focus is given on the method and results obtained with the ANN, not on the ANN.•Calculations performed over a vented explosion database containing 268 tests.•A data augmentation method based on physical measurements is proposed.•Routes are proposed to perform relevant analyses on the ANN predictions.