Assessment of machine learning algorithms for predicting autoignition and ignition delay time in microscale supercritical water oxidation process

Abstract

With recent advancements in space technology, there is a need to develop technologies to ensure a sustainable environment for human survival. Among these, treatment of human and organic waste aboard manned space missions is a challenging task for which supercritical water oxidation using hydrothermal flames has been proposed as a possible solution. The critical step in readily adopting this technology from established ground-based setups is scaling the process to microscale. In addition to the challenge of physical realization of the microreactors at these high pressure and temperature (P > 22 MPa, T>350°C) conditions, the need to explicitly analyze the process dynamics at microscale is inevitable owed to the size of the reactors under consideration, the physics being significantly different from meso/mini scale systems. One of the primary objectives is to identify the operating physical parameters for which formation of hydrothermal flames can be obtained. Before proceeding with an expensive computational or experimental approach to determine the exact ignition map, an initial estimate based on physical arguments can help in providing insights into the process. We address this problem using homogeneous ignition calculations to develop machine learning models to predict autoignition as well as ignition delay time. The ingenuity of the work lies in defining autoignition criteria in relation to flow time scales expected at microscale. Various classification models were trained and tested for predicting autoignition and regression models were demonstrated to predict the ignition delay time. While predicting autoignition is a straightforward process, a two-step approach is proposed for ignition delay time. Finally, how machine learning can be used more explicitly, particularly for understanding and designing efficient microreactors, is presented which highlights that machine learning approach is not merely restricted to prediction but can also have real implications on improving the process as a whole.