Prediction of blast loads using machine learning approaches

The assessment of human injuries and structural damage following the detonation of a high explosive requires an understanding of blast load parameters. Use of physical experiments or physics-based numerical tools require large amounts of time and expertise, often restricting their use to deterministic analyses. Since explosive events are inherently unpredictable and key variables (e.g. charge size, mass, composition, location) may not be known a priori, there is a clear need for rapid analysis tools that can embrace this uncertainty in a probabilistic framework. Machine learning tools have been developed for this purpose, however, the features of the problem that are selected as model inputs can result in predictions being fixed to a single domain, thus requiring the tool to be retained for every new scenario. This paper details how the Directionencoded Neural Network (DeNN), a novel Machine Learning method, takes inspiration from the operation of robot vacuum cleaners to prevent this issue by considering the surroundings of each prediction point. Through comparisons to a traditional Artificial Neural Network (ANN), provided with global domain inputs, it is shown that the DeNN’s unique feature selection process allows for predictions in domains of variable sizes with movable obstacles, ultimately producing a tool that can be used in a range of studies without requiring additional task-specific training.