UC Berkeley

This dissertation presents a systematic investigation of machine learning strategies for analyzing boiling heat transfer on zinc oxide (ZnO) nanostructured surfaces, progressing from single-modality digital data analysis to multimodal sensor fusion incorporating image, digital, and acoustic data. ZnO nanopillar surfaces with ultra-low contact angles below 10 degrees exhibit distinct vaporization regimes during water droplet deposition experiments at surface superheats ranging from 10°C to 55°C.The research advances through three phases. First, three convolutional neural network (CNN) architectures with progressively more complex input modalities are developed for combined image and sensor data analysis. A digital-only baseline (Case A) achieves 15.8% root mean squared percent error (RMSPE) in mean heat flux prediction. A hybrid parallel-series CNN incorporating high-speed video images with non-thermal digital data (Case B) reduces this to 10.3% RMSPE, demonstrating that the CNN effectively functions as a non-contact temperature sensor by extracting thermal state information from visual morphology without direct temperature measurement. Adding all available digital inputs through a skip connection architecture (Case C) achieves 5.8% RMSPE and 96.9% regime classification accuracy. Second, acoustic emission monitoring is introduced as an additional sensing modality. A neural network trained on quartile-based acoustic features (mean absolute amplitude, mean frequency, and frequency standard deviation computed over four equal temporal segments of each signal) achieves 13.3% RMSPE in predicting surface superheat from sound alone, demonstrating that acoustic emissions carry quantitative information about bubble nucleation and collapse dynamics complementary to visual and thermal measurements. Third, a multimodal fusion network incorporating image, digital, and acoustic data achieves 4.6% RMSPE in heat flux prediction and 100% boiling regime classification accuracy. Permutation importance analysis confirms that image and acoustic features contribute most, with acoustic information providing its greatest advantage in the vigorous nucleation and Partial Dryout Vapor Recoil (PDVR) regimes. The progressive improvement from 15.8% to 4.6% RMSPE provides compelling evidence that multimodal sensor fusion represents a powerful paradigm for analyzing complex thermal-fluid systems, with broad applicability to flow boiling, condensation, and nuclear thermal-hydraulics.