Using Machine Learning for Explosion Yield Estimation

Abstract

ABSTRACT We developed a machine learning approach to estimate an explosion yield and height-of-burst or depth-of-burial (HOB or DOB) using a combination of seismic ground-motion and acoustic measurements. The technique employs artificial neural networks (ANNs) with Bayesian regularization suitable for small datasets to reduce the potential for overfitting and improve the network generalization. Using data from multiple explosion experiments conducted in various rock types, we investigated the effect of different seismic and acoustic measurement methods combined with estimated seismic velocity on the yield and the scaled HOB or DOB estimates. The training dataset for the ANN comprised data from 42 explosions conducted in various media at different HOB or DOB. Two additional explosions were used for the method validation. The input features included seismic and acoustic amplitudes, peak frequency, positive phase duration, and the apparent seismic velocity. The presented approach is not limited to a single lithology. Instead, it uses a diverse set of parameters, such as apparent seismic velocity and P-wave peak frequency, to implicitly identify the medium properties within the data-driven framework. Incorporating these parameters resulted in significant improvement of the method performance for both scaled depth and yield estimates. The root mean square error of the scaled depth estimate is on the order of 0.1 m/kg1/3. For the yield estimate, the mean absolute percentage error is less than 10% for both training and validation datasets. The main challenge of using machine learning for yield estimate is a small number of calibration explosions with known yields available for the ANN training. In the future, the developed approach can be further improved by training the ANNs with larger datasets, as more explosion data become available.