Semi-surrogate modelling of droplets evaporation process via XGBoost integrated CFD simulations

Yihuan Yan; Xueren Li; Weijie Sun; Xiang Fang; Fajiang He; Jiyuan Tu

doi:10.1016/j.scitotenv.2023.164968

Semi-surrogate modelling of droplets evaporation process via XGBoost integrated CFD simulations

Sci Total Environ. 2023 Oct 15:895:164968. doi: 10.1016/j.scitotenv.2023.164968. Epub 2023 Jun 24.

Authors

Yihuan Yan¹, Xueren Li², Weijie Sun³, Xiang Fang⁴, Fajiang He⁴, Jiyuan Tu²

Affiliations

¹ School of Air Transportation / Flying, Shanghai University of Engineering Science, Shanghai 201620, China. Electronic address: yihuan.yan@outlook.com.
² School of Engineering, RMIT Unversity, PO Box 71, Bundoora, VIC 3083, Australia.
³ Department of Computing Science, University of Alberta, Edmonton, AB T6G 2R3, Canada.
⁴ School of Air Transportation / Flying, Shanghai University of Engineering Science, Shanghai 201620, China.

PMID: 37356762
DOI: 10.1016/j.scitotenv.2023.164968

Abstract

The applications of machine learning (ML) based approach are emerging as possible tools to accelerate CFD simulations. This study proposed a semi-surrogate model for CFD with integration of the cutting-edge ML algorithm, eXtreme Gradient Boosting (XGB), which enlightened a possible pathway to effectively and efficiently solve and predict those costly but highly repetitive fluid dynamics-related problems. Droplet evaporation, a complex but essential phenomenon in respiratory droplets transport, was studied as the practical case using the proposed model. Droplets evaporation and dynamic size distributions were firstly tracked under various combinations of indoor humidity and temperature using traditional Eulerian-Lagrangian CFD framework, followed by generating several datasets for XGB training. The trained XGB was then used to interpret the evaporated droplets size over time under new combinations of indoor conditions. Outcomes revealed that well-trained XGB-base semi-surrogate model was capable of interpreting complex non-linear relationships between droplets dynamic parameters (diameter and time) and indoor parameters (humidity and temperature). For each specific parameter, the predictive error of well-trained XGB could retain below 5 % and its prediction speed was found nearly 1 million times faster than that of new CFD simulations. Successful applications of XGB in conjunction with CFD demonstrated its great potential on providing rapid and more efficient predictions of complex, costly and repetitive fluid dynamics-related phenomenons (e.g. droplets evaporation). Also, the XGB predicted droplets evaporation data from this study could be further applied as initial conditions into new simulations via the User-defined function (UDF).

Keywords: CFD; Droplets evaporation; Semi-surrogate model; Supervised machine learning algorithm; eXtreme gradient boosting (XGB).