posted on 2023-09-06, 17:37authored byJiguo Tang, Shengzhi Yu, Xiaofan Hou, Tianhui Wu, Hongtao Liu
Drop impact on a solid surface is a fundamental phenomenon
in nature
and engineering. Prediction of the maximum spreading ratio during
drop impact is critical for modeling and optimizing the relevant processes.
However, accurately modeling the maximum spreading using empirical
and numerical methods remains challenging. Machine learning (ML) has
recently provided a promising way to understand and model complex
fluid phenomena. Thus, in this study, a universal model is developed
by using machine learning methods to predict the maximum spreading.
TPE (Tree-Structured Parzen Estimator) algorithm, a variant of Bayesian
optimization, is applied to optimize the hyperparameters to improve
the predictive performance of ML models. An extensive database containing
1015 experimental data points has been constructed from 24 research
sources and the present experimental results. Four boosting ML models
were compared with conventional models, and the results show that
the mean absolute percentage error (MAPE) of ML models is 2.62–3.40%,
which is less than a third that of the best conventional model. Among
these ML models, the TPE-based CatBoost model is superior to others,
with an MAPE of 1.67% and an R2 of 0.952.
Then, SHAP (SHapley Additive exPlanations) was used to address the
black-box nature of the ML-based models. Parameter analysis indicates
that the developed ML model robustly captures the physical variation
trend of the maximum spreading for various working fluids. The results
presented here demonstrate that the TPE-based CatBoost model can model
the drop maximum spreading ratio with high accuracy and broad applicability.