Prospective on applying machine learning in computational fluid dynamics (CFD) simulation of metallurgical reactors

Abstract

Metallurgical reactors, especially in ironmaking/steelmaking process, characterise with high-temperature turbulence, multiphase flow, mass/heat transfer and reactions. Computational fluid dynamics (CFD) simulation-based design and optimisation are of significance for efficient metallurgical performance. However, the difficulty and cost to numerically solve the nonlinear controlling equations combined with data pre/post-processing make the whole CFD simulation process time-consuming, which makes it challenging to provide in-time feedback for industrial practices. The popularisation and prosperous development of machine learning bring new opportunities for promoting CFD performance. Discussion has been made on the current research progress of applying machine learning in the whole CFD workflow including pre-processing, solving, and post-processing. Among them, the time consumed by manual pre-processing exceeds 50% of CFD tasks in general. The machine learning or parametric modelling methods can reduce pre-processing time by three orders in the estimate. The solving step is expected to be accelerated by 5 to 1000 times using machine learning. A brief review of machine learning coupled CFD is provided, as is a prospective on its development. Discussion is presented on the main functions, challenges, typical techniques and future directions of applying machine learning in CFD simulation of metallurgical reactors, for the purpose of making CFD faster, more accurate, and better visualised based on the metallurgical practices.