Machine Learning Enabled Surface Temperature, Heat Flux, and Phase Field Data Reconstruction Boiling

Abstract 

Prediction of boiling is critical for designing efficient thermal management systems that leverage liquid-to-vapor phase change to address the needs of energy-intensive applications, including cooling of AI computing infrastructure and electrified transportation systems.  However, traditional correlations and models are typically developed and validated using spatially averaged measurement data, despite observations of complex spatiotemporal phenomena and awareness of the multiple constituent mechanisms contributing to boiling heat transfer. Emerging predictive methods for boiling transport propose to, instead, develop or train predictive models leveraging multi-modal datasets that characterize thermal and phase field information at high spatiotemporal resolution. This talk will summarize our recent efforts in the acquisition and interpretation of synchronized through-substrate visualizations of the wall-adjacent liquid-vapor phase distribution and temperature of boiling surfaces, along with side-view visualizations, during flow boiling of low-surface-tension dielectric fluids.  Machine learning techniques are explored as tools to further enable this metrology. Namely physics-informed neural networks are leveraged to robustly reconstruct surface temperatures and heat fluxes from noisy measurement data. Deep learning neural operators are then trained on these data as synthetic cameras for real-time inference of heat flux and phase fields from infrared images. And lastly, generative frameworks are trained to infer surface temperature maps beyond the infrared field-of-view. As part of a broader multi-disciplinary collaboration, these high-fidelity spatiotemporal data at the boiling surface, combined with optical metrologies being developed in parallel to probe interfaces and two-phase velocity and temperature fields in the bulk liquid, offer a path towards robust validation of flow boiling models and training of surrogates for predicting boiling at full complexity, utilizing a combination of physics-informed and data-driven machine learning methods.

 About the Speaker 

Justin A. Weibel is a Professor of the School of Mechanical Engineering at Purdue University, as well as Director of the Cooling Technologies Research Center (CTRC), a graduated NSF I/UCRC that addresses research and development needs of companies and organizations in the area of high-performance heat removal from compact spaces. His research group explores methodologies for prediction and control of heat transport to enhance the performance and efficiency of thermal management technologies. He has been a key contributor to the development of transformative technologies supported by the DARPA TGP (2008-2012), DARPA ICECool (2013-2017), NAVSEA NEEC (2016-2018), ONR NEPTUNE (2015-2021), SRC CHIRP (2019-2025), ARPA-E ASCEND (2021-2025), and ARPA-E COOLERCHIPS (2023-) programs, in addition to numerous sponsored research projects that transition these technologies to industry. He PI of the METHODS (Machine learning Enabled Two-pHase flow metrologies, models, and Optimized DesignS) ONR MURI program (2024-2029). Dr. Weibel has supervised 46 PhD and 16 MS students, current and former, co-authored over 250 refereed journal and conference papers (h-index of 52). He has been recognized as an Outstanding Engineering Teacher and Outstanding Faculty Mentor in the College of Engineering at Purdue University and is a University Faculty Scholar (2025-2030). He received the 2020 ASME Electronic & Photonic Packaging Division (EPPD) Young Engineer Award, 2021 ASME K-16 Outstanding Early Faculty Career in Thermal Management Award, and the 2024 ASME K-16 Clock Award; in 2023 he was elected a Fellow of the ASME. Dr. Weibel is on the IEEE ITherm Executive Committee and is Associate Editor of the IEEE Transactions on Components Packaging and Manufacturing Technology and ASME Journal of Electronics Packaging.

Host: Professor Nenad Miljkovic