Calculating the view factor of randomly dispersed multi-sized particles using hybrid GRU-LSTM recurrent neural networks regression

Abstract

This study presents three novel methods for the computational acceleration of simulations of high temperature radiative heat transfer in particulate media. The paper presents a novel algorithm for neighbor searching that can locate and sort multi-sized particles, based on their distance, with minimal computation. The traditional use of regression recurrent neural networks for time series data is modified to a regression function for distance-ordered data. LSTM and GRU cells are used in a deep recurrent neural network to predict radiation view factors of particle and face neighbors through the distance-ordered sequential data. The preprocessing of sequential data is addressed with two new transformation methods. Through the use of the transformation methods, both the particle and face neighbors are regarded as general objects, allowing one regression neural network to predict the view factor of any number of particles and faces sequentially, which represents a unique feature. With neural networks and preprocessing methods, it is possible to predict the view factor of different-sized objects (particles and faces), representing a significant development in the field. The models demonstrate high precision for objects located within the short-range radiation region. Objects located a long distance from the emitting particle have an over-prediction of view factor, but due to their negligible contribution to overall radiation heat transfer, a high coefficient of determination is obtained for particle and face neighbors, with several orders of magnitude improvement in computational speed compared to typical Monte Carlo ray tracing methods.