Learning pool boiling physics with scientific machine learning

相变是支撑许多自然现象和工业应用的基本过程，如昆虫的露水凝结，水滴收集，电子设备的浸入式冷却和核反应堆冷却。沸腾是一种特殊的相变过程，它涉及气泡的动态形成、运动和演化。气泡标志着液体-蒸汽界面，它们的统计特征对于揭示基本的沸腾原理至关重要。到目前为止，由于气泡非常复杂、运动迅速、变形剧烈，因此表征和提取气泡统计数据非常具有挑战性。这些关于目标跟踪和数据处理的限制可以通过计算机视觉和深度学习的最新进展来解决。在这项提出的工作中，计算机模型被训练来检测和跟踪实时成像数据中的气泡，并随后提取有关气泡的有意义的统计数据，例如，它们的数量，大小和轨迹，以学习气泡和沸腾物理。在热流体工程中使用计算机视觉概念的成功将通过集成工程、计算机科学和数据科学中的新技术，使表面设计、气泡数据和沸腾性能之间的相互关系建模成为可能。该项目的目标是开发一种新的基于数据的方法来获得对沸腾过程的新理解。该提案设想了三个主要重点：(1)通过实验和计算从实时气泡图像中提取可解释且丰富的物理描述符，这在过去一直是一个挑战；(2)通过迁移学习、概率代理建模和序列学习，建立连接表面、气泡和热性能的学习模型；(3)提供对动态沸腾物理的整体和基本理解，使具有所需传热性能的表面逆设计成为可能。另一个贡献将是用机器学习、数据科学和统计学的概念和方法丰富热科学界。这种富集通常会对许多热流体实验产生积极影响，包括生物流体中的细胞或颗粒跟踪，冷凝过程中的液滴研究以及流动沸腾过程中的气泡流统计。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Phase changes are fundamental processes that underpin many natural phenomena and industrial applications such as dew condensation on insects, water droplet harvesting, immersion cooling for electronic devices, and nuclear reactor cooling. Boiling is a particular phase change process which involves the dynamic formation, movement, and evolution of bubbles. Bubbles mark liquid-vapor interfaces, and characterization of their statistics is crucial for unveiling fundamental boiling principles. So far, it has been very challenging to characterize and extract bubble statistics as they are extremely complex, rapidly move, and deform drastically. Such limitations about object tracking and data processing can be addressed by the recent advances in computer vision and deep learning. In this proposed work, computer models are trained to detect and track bubbles in live-imaging data and, subsequently, extract meaningful statistics on bubbles, for example, their number, size, and trajectory, to learn bubbles and boiling physics. The success of using computer vision concepts in thermofluidic engineering will enable modelling the interconnected relationships among surface designs, bubble data, and boiling performances by integrating new technologies in engineering, computer science, and data science. The goal of this project is to develop a novel data-based approach to obtain new understanding about boiling processes. The proposal envisions three major thrusts: (1) Experimentally and computationally extract interpretable and rich physical descriptors from live bubble images, which has been challenging in past; (2) develop learning models to connect surfaces, bubbles, and thermal performances by applying transfer learning, probabilistic surrogate modelling, and sequence learning; and (3) provide a holistic and fundamental understanding of dynamic boiling physics, enabling the inverse design of surfaces with desired heat transfer performance. Another contribution will be enriching the thermal science community with concepts and methods from machine learning, data science, and statistics. This enrichment can positively impact numerous thermofluidic experiments, in general, including cell or particle tracking in biofluidics, droplet study during condensation, and bubbly flow statistics during flow boiling.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.