Characterization and Control of Crack Propagation and Coalescence

Abstract

Four stochastic inversion and optimization methods (MCMC, SA, PSO, and ACO) are investigated and compared by coupling with a novel mechanistic model for subsurface characterization to jointly estimate the subsurface properties. For the first time, the newly developed log-based subsurface parameter estimation technique can continuously estimate in-situ wettability (contact angle) along with water conductivity, clay surface conductance, and oil saturation as a function of depth along the wellbore. The proposed methods have significance for various electromagnetic-based characterization of fluid-filled porous geological materials.

This is followed by an investigation of the DDPG deep reinforcement learning algorithm on crack propagation/growth control, where the RL agent in the form of neural network is trained and optimized by continuously interacting with the environment to obtain an optimal control policy. It starts from a simple problem to control the direction of crack propagation in a self-defined environment and extends to a more complex problem to control both the direction and rate of mixed-mode fatigue crack growth. The RL agent is able to develop an optimal and computationally tractable control policy that adaptively changes the engineering parameters in both problems. The potential application of the control framework to the real-world environment is also demonstrated. Most importantly, a first-of-its-kind design of reward function through reward shaping for a generalizable problem of controlling spatiotemporally propagating processes is presented, which helps the RL agent to quickly and correctly converge to the optimal policy.

In the last part of this dissertation, the crack coalescence is characterized with the help of machine learning technique and the novel FDEM-based simulator HOSS. The models representing a 2D rock-like material containing two parallel inclined initial slots with various placements under uniaxial compression is built in Cubit, and the simulation data is generated using HOSS. The features that are indicators of the underlying mechanism of crack propagation and coalescence are extracted from the simulation results as inputs of a SVM classifier with RBF kernel. Through permutation feature importance evaluation, the top 7 most important features are identified, from which the status of energy buildup, distribution and release inside the material can be inferred.