Monitoring and Controlling Mechanical Discontinuity Using Supervised Learning Coupled with Reinforcement Learning

The ability to accurately monitor and subsequently control mechanical discontinuity is a grand challenge posed by the U.S. Department of Energy and National Science Foundation. We present a reinforcement learning framework that interacts with a mechanistic equation or numerical simulator to learn to control the propagation of mechanical discontinuity. The deep deterministic policy gradient (DDPG) algorithm is implemented for learning continuous actions for controlling the propagation of mechanical discontinuity. The reward function that forces the propagation along the shortest linear path between tip of the discontinuity and the final goal point performs the best. The Markov decision process, which includes state, action and reward, had to be carefully designed to obtain a good control policy. The above-mentioned control of fracture propagation depends on accurate detection and monitoring of mechanical discontinuity. We developed a generalized supervised-learning framework for visualizing the evolution of single, linear mechanical discontinuity embedded in materials of any composition, velocity, density, porosity, and size with limited data. The proposed visualization of discontinuity requires accurate estimations of the length, location, and orientation of the propagating discontinuity by processing waveforms recorded during multi-point wave-transmission measurements. The approach considers the wave attenuation, dispersion, and mode conversion. Statistical feature extraction, gaussian random process modeling, and singular value decomposition are shown to be essential data preprocessing steps. For purposes of monitoring, a double-layered neural network-based regression has the best performance with an accuracy of 0.95 to 0.97, in terms of R2, within a 95% confidence interval. The performance drops less than 0.03 in R2 with the drop in the precision of measurements from 0.01 Pa to 1 Pa and drop in sampling rate from 10 MHz to 1 MHz in addition to the presence of 10% Gaussian noise. An accurate monitoring of the propagation allows robust control of the propagation of mechanical discontinuity. Robust control is achieved in 500 learning episodes involving at the maximum 25 iterations each episode.