Development of a custom supervised learning network to capture high-temperature oxidation of carbon fibers

Carbon fibers form the basis of ablative heat shield materials that are used in hypersonic flight. Detailed thermochemical flow simulations and experimental characterization has provided valuable insights into the oxidation of individual carbon fibers at high temperatures. However, flow simulations that resolve individual fibers are expensive and cannot be easily integrated into macroscale solvers. In this work, we propose a supervised learning model capable of capturing the spatio-temporal evolution of carbon fibers that are oxidized by oxygen atoms. The proposed supervised learning model is inspired by conventional neural networks used for temporal image-based prediction. The training microstructural images consists of snapshots depicting the physical changes in a single cylindrical fiber that are obtained through Direct Simulation Monte Carlo (DSMC) simulations in conjunction with a synthetic microstructure generation code called Fibergen, where snapshots of a single cylindrical fiber subject to oxidation are recorded. A user defined surface function acts as the encoder and decoder of the proposed supervised learning model. Support Vector Regression (SVR) forms the heart of the proposed predictive model by relating the user defined surface function parameters with input variables (time and microscopic probabilities). The developed supervised learning model is capable of resolving the two distinctive types of oxidation, needling and thinning, which is observed when a fiber undergoes oxidation with varying microscopic probabilities. Once trained, the developed supervised learning model is able to predict the various states of oxidation of a single fiber in a very short period (few seconds) as opposed to expensive DSMC simulations. The final model consists of seven analytical functions directly dependent on the input variables which facilitates the smooth integration of the supervised learning model with the macroscale solvers. In other words, the proposed supervised learning model provides a novel and an effective pathway to relate the microscale changes in carbon microstructures with macroscale solvers.