Predicting flow-blurring droplet size using neural networks and Bayesian optimization: A data-driven approach

Abstract

Flow-blurring injectors, known for producing fine sprays in twin-fluid systems, are essential for applications involving high-viscosity fuels, such as biofuels. This study presents a data-driven approach using neural networks and Bayesian optimization to predict the Sauter Mean Diameter (SMD) of flow-blurring sprays. A dataset from the experimental literature was curated and pre-processed, with critical dimensionless parameters – including the Reynolds number, Weber number, injector’s aspect ratio, and air-to-liquid mass flow rate – used to train multi-layer perceptron (MLP) models. Through Bayesian optimization, hyperparameters such as neuron count, learning rate, and regularization were fine-tuned to enhance model accuracy and avoid overfitting. The optimized models achieved high predictive accuracy, with regression scores exceeding 97% and minimal mean-squared error (MSE), demonstrating that Bayesian-optimized neural networks can significantly reduce reliance on costly experimental and numerical methods. This approach provides a fast, accurate solution for spray modeling, offering a scalable method for optimizing injector designs in fuel systems, particularly for alternative fuel applications.