Modelling flame-to-fuel heat transfer by deep learning and fire images

Caiyi Xiong,Zilong Wang,Xinyan Huang

DOI: https://doi.org/10.1080/19942060.2024.2331114

IF: 6.519

2024-03-22

Engineering Applications of Computational Fluid Mechanics

Abstract:In numerical fire simulations, the calculation of thermal feedback from the flame to the solid and liquid fuel surface plays a critical role as it connects the fundamental gas-phase flame burning and condensed-phase fuel gasification. However, it is a computationally intensive task in CFD fire modelling methods because of the requirement of a high-resolution grid for calculating the interface heat transfer. This paper proposed a real-time prediction of the flame-to-fuel heat transfer by using simulated flame images and a computer-vision deep learning method. Different methanol pool fires were selected to produce the image database for training the model. As the pool diameters increase from 20 to 40 cm, the dominant flame-to-fuel heat transfer shifts from convection to radiation. Results show that the proposed AI algorithm trained by flame images can predict both the convective and radiative heat flux distributions on the condensed fuel surface with a relative error below 20%, based on the input of real-time flame morphology that can be captured by a larger grid size. Regardless of growing or decaying fires or puffing flames induced by buoyancy, this method can further predict the non-uniform distribution of heat transfer coefficient on the interface rather than using empirical correlations. This work demonstrates the use of AI and computer vision in accelerating numerical fire simulation, which helps simulate complex fire behaviours with simpler models and smaller computational costs. Highlights

engineering, mechanical,mechanics, multidisciplinary

What problem does this paper attempt to address?

The paper aims to address the real-time prediction of heat transfer between flames and fuel in fire simulations. Specifically: 1. **Research Background**: In fire simulations, heat transfer from the flame to the fuel surface (including convection and radiation) is a key factor because it connects the processes of gas-phase combustion and condensed-phase fuel vaporization. However, in computational fluid dynamics (CFD) fire simulation methods, accurately calculating interface heat transfer requires high-resolution grids, which leads to very high computational costs. 2. **Core Issue**: How to use computer vision and deep learning technologies to predict the heat transfer from the flame to the fuel surface in real-time, thereby simplifying the computational model and reducing computational costs. Traditional methods rely on empirical correlations, which are usually not based on actual fire tests and cannot capture the dynamic changes of turbulent flames. 3. **Solution**: The paper proposes a method that uses simulated flame images and computer vision deep learning techniques to predict the heat transfer coefficient from the flame to the fuel surface in real-time. This method can significantly reduce the dependence on fine grids, allowing the use of coarse grids to accurately simulate complex fire phenomena, thereby reducing computational costs. 4. **Experimental Validation**: Through simulation experiments of methanol pool fires with different diameters, the proposed deep learning model was validated to predict the distribution of convective and radiative heat fluxes with a relative error of less than 20%. Additionally, this method can capture the non-uniform distribution of heat transfer coefficients and is applicable to various fire scenarios, such as flame pulsation, flame expansion, or contraction. In summary, this paper aims to accelerate the fire simulation process and improve the accuracy and efficiency of simulations through artificial intelligence and computer vision technologies.

Modelling flame-to-fuel heat transfer by deep learning and fire images

Heat Flux Prediction of Radiation Balance Wall by Deep Convolutional Neural Networks

DeepFlame: A deep learning empowered open-source platform for reacting flow simulations

Flame Detection Using Deep Learning

Flame Detection in Electric Power Scenarios Based on Gaussian Modeling

GPU-accelerated Large Eddy Simulation of turbulent stratified flames with machine learning chemistry

Deep Learning-Driven Analysis for Cellular Structure Characteristics of Spherical Premixed Hydrogen-Air Flames

A deep learning-based dynamic deformable adaptive framework for locating the root region of the dynamic flames

A Priori Assessment of Convolutional Neural Network and Algebraic Models for Flame Surface Density of High Karlovitz Premixed Flames

Using machine learning in physics-based simulation of fire

Flow and thermal radiation characteristics of a turbulent flame by large eddy simulation

Predictive Models for Flame Evolution Using Machine Learning: A Priori Assessment in Turbulent Flames Without and with Mean Shear

Predicting Spatiotemporal Distributions in a Bubbling Fluidized Bed for Biomass Fast Pyrolysis Using Convolutional Neural Networks

Graphics Processing Unit/artificial Neural Network-Accelerated Large-Eddy Simulation of Swirling Premixed Flames

Modeling of Combustion Characteristics of Particles in Transient Gas–Solid Reacting Flow Via a Machine Learning Approach

Meshed axisymmetric flame simulation and temperature reconstruction using light field camera

Modelling turbulent premixed flames using convolutional neural networks: application to sub-grid scale variance and filtered reaction rate

Advanced Image Processing Techniques for Multi-Level Characterization of Significant Flame Features in Carbon-Neutral Combustion

Prediction of Flame Type and Liftoff Height of Fuel Jets in Turbulent Hot Coflow Using Machine Learning Methods

High-dimensional FGM-ResNet modelling of turbulent spray combustion: Effects of evaporation non-adiabacity and scalar correlation

Graphics Processing Unit/Artificial Neural Network-accelerated large-eddy simulation of turbulent combustion: Application to swirling premixed flames