Machine learning enabled condensation heat transfer measurement

Abstract

Measuring condensation heat transfer and its associated heat transfer coefficient is not trivial. Rigorous measurements require careful experimental design and tradeoff studies to properly select sensor type, sample geometry and size, coolant fluid and flow rate, operating conditions, working fluid purity, purge methodology, and measurement protocol. Conventional tube-based condensation heat transfer measurements quantify the change in the enthalpy of a single-phase coolant flow via measurement of the inlet and outlet bulk coolant temperatures. The uncertainties associated with this classical and well-established experimental method are high. The high uncertainty stems from the high characteristic heat transfer coefficient or heat flux associated with the condensation process, making the thermal resistance on the external tube side typically on the same order of magnitude as the internal single-phase coolant convective heat transfer thermal resistance. Even when taking the utmost care and using extremely accurate sensors having low uncertainty, the relative uncertainties of heat flux and heat transfer coefficient can be in the range of 20% to 100%. Here, we take advantage of machine learning (ML) to develop an optical visualization method for dropwise condensation heat transfer characterization. Using state-of-the-art intelligent vision, we demonstrate a previously-unexplored method for characterizing the condensate droplet shedding frequency, droplet shedding size, and heat flux without the need for high-speed imaging. We verify our technique by conducting rigorous steam condensation measurements on Parylene C coated smooth copper tube samples having 500 nm, 1 μm, and 5 μm Parylene C thicknesses. We validate our ML predictions with data obtained simultaneously using the enthalpy-change method on a custom and well-established condensation chamber. In contrast to conventional heat transfer measurement methods, the uncertainty of our ML method is constant (∼10%) and does not vary with heat flux. We finally demonstrate the key advantage of our ML measurement technique on a custom-made tube having axially varying surface properties resulting in differing local heat transfer coefficient. Our ML heat transfer measurement method enables the high fidelity characterization of phase change heat flux, reduction in relative measurement uncertainty, resolution of local effects, and elimination of the need for temperature measurement across samples.