Accurate machine-learning-based prediction of aerodynamic and heat transfer coefficients for cylindrical biomass particles

Abstract

Cylindrical biomass (straw) particles play a crucial role in the numerical simulations of biomass co-firing in coal-fired boilers, wherein the aerodynamic and heat transfer coefficients of these particles are particularly important. Despite the existence of models proposed thus far for various specific conditions, developing a universal model for cylindrical particles remains challenging in multiphase flow systems. In this paper, a total of 1092 validated computational fluid dynamics (CFD) datasets of cylindrical drag, lift, torque coefficients and average Nusselt number are acquired based on direct numerical simulations of OpenFOAM body-fitted meshes, and a hybrid algorithmic framework of N+1 convolutional neural network (CNN) machine learning optimality is established. where multiple single-tree models (referred to as N models) are combined with an additional CNN layer (referred to as +1 model).The framework inherits the advantages of a single-tree models while significantly improving the overall algorithmic generalisation performance, with the aim of allowing efficient use of time and minimal expertise in the fluid modelling process. The mean relative absolute errors for drag, lift, torque coefficient, and the average Nusselt number was evaluated, and the results demonstrated a reduction of 75.19 %, 89.48 %, 89.53 %, and 85.70 %, respectively, in the mean relative absolute errors compared to the extremely randomised tree single-tree model. We extended the model scope to different aspect ratios (1 ≤ Ar ≤ 30), angles of incidence (0° ≤ θ ≤ 90°), and Reynolds numbers (1 ≤ Re ≤ 4000), and conducted random predictions. Compared with traditional mathematical relationships, the proposed framework reduced the average relative error to 0.99 %. The simultaneous application of models for the prediction of unknown parameters was validated against relevant early literature, which demonstrated that the average relative error of drag coefficients for the predicted values remained within 2.07 %. The N+1 (CNN) hybrid algorithm framework exhibited high accuracy and excellent adaptability, furnishing a substantial quantity of data to support the Euler–Lagrange model of non-spherical biomass particle multiphase flow in industrial applications.