OPTIMIZING HYPERPARAMETER TUNING IN MACHINE LEARNING MODELS FOR FLOW BOILING IN MICROTUBES WITH BAYESIAN INFORMATION CRITERION

Abstract

Flow boiling in microtubes is a promising thermal management technology because microtubes can deliver very high heat transfer rates and accommodate heat fluxes beyond the limits of conventional cooling methods. Despite this potential, widespread application of such technology on a commercial scale is still challenging. One of the main challenges is associated with the lack of reliable predictive models capable of capturing the impact on the heat transfer coefficient of various geometries, materials, and operating conditions. To tackle this challenge, this work uses a data-driven modelling framework using artificial neural networks (ANNs) integrated with Bayesian Optimization (BO) techniques for hyperparameter tuning. Two objective functions are compared: mean absolute error (MAE), which emphasizes predictive accuracy, and Bayesian Information Criterion (BIC), which provides a balance between model accuracy and complexity to avoid underfitting or overfitting. To develop the models, a preliminary analysis of the flow boiling data is carried out to remove unwanted data and outliers. Therefore, the models are developed and optimized for the prediction of the flow boiling heat transfer coefficient (HTC). Prediction results show that MAE-based optimization favors a more complex ANN and achieves lower test-set errors (MAE ≈ 261 W/m²·K, MAPE ≈ 3.1%) than the BIC-optimized model (MAE ≈ 387 W/m²·K, MAPE ≈ 5.4%), while the BIC model uses substantially fewer parameters. However, this improvement is only 1.1% with respect to the average heat transfer coefficient value of the dataset. Overall, the study illustrates the trade-off between accuracy and simplicity and demonstrates the effectiveness of BO-enabled ANN frameworks for HTC prediction in microchannel flow boiling.