物理集成神经网络建模：静态与动态工况下低温液体存储中的热质传递
Physics-integrated neural network modeling of heat and mass transfer in cryogenic liquid storage under static and dynamic conditions
- von Karman Institute for Fluid Dynamics（冯·卡门流体力学研究所）
- Université Libre de Bruxelles（布鲁塞尔自由大学）
- Universidad Carlos III de Madrid（马德里卡洛斯三世大学）
- École Polytechnique de Bruxelles, Université Libre de Bruxelles（布鲁塞尔皇家理工学院，布鲁塞尔自由大学）
机构由 AI 辅助整理，请以论文原文为准。
AI总结：
该研究提出物理集成神经网络框架，结合守恒型零维节点模型与数据驱动封闭项，经48次低温储罐实验验证，全局归一化均方根误差低于3%，可高效提取物理解释性封闭律用于低温液体存储热质传递预测。
AI中文摘要：
低温液体存储的系统级精准预测仍具挑战性，因为降阶模型依赖于与工况相关的封闭项来处理未 resolvable 的热质传递，尤其是在晃荡工况下。本文提出一种物理集成神经网络框架，该框架将基于守恒律的零维节点模型与数据驱动的封闭项相结合，用于壁面与界面热传递、相变、增压剂品质以及液体热边界层演化。液体内部的热分层与混合通过边界层厚度的一阶动力学模型表征，而四个针对不同工况的神经网络则推断出自增压与弛豫、主动增压、排气以及横向晃荡的封闭参数。该框架利用在光学可观测设施中开展的48次多阶段低温储罐实验的专用数据库进行识别与评估，实验采用液氮开展，涵盖受控壁面加热、蒸汽注入与抽排、强制横向晃荡，既包含缓慢演化的热态，也包含强瞬变工况。通过实验级K折交叉验证评估泛化能力，损失函数中加入熵产生惩罚项以避免违反热力学第二定律。在所研究的交叉验证配置中，全局归一化均方根误差始终低于3%，其中晃荡工况是最具挑战性的工况。在完整数据库上训练的最终模型以1.6%的全局误差重现了实验结果。这些结果表明，所提框架可从有限但信息丰富的实验数据库中提取具有物理解释性且计算高效的封闭律。
英文摘要：
Accurate system-level prediction of cryogenic liquid storage remains challenging because reduced-order models rely on regime-dependent closures for unresolved heat and mass transfer, particularly under sloshing. We present a physics-integrated neural-network framework that combines a conservation-based zero-dimensional nodal model with data-driven closures for wall and interfacial heat transfer, phase change, pressurant quality, and liquid thermal-boundary-layer evolution. Thermal stratification and mixing within the liquid are represented through a first-order dynamical model for the boundary-layer thickness, while four operating-regime-specific neural networks infer the closure parameters for self-pressurization and relaxation, active pressurization, venting, and lateral sloshing. The framework was identified and evaluated using a dedicated database of 48 multi-stage cryogenic-tank experiments conducted in an optically accessible facility operated with liquid nitrogen. The experiments combine controlled wall heating, vapor injection and evacuation, and forced lateral sloshing, thereby covering both slowly evolving thermal states and strongly transient operating conditions. Generalization was assessed through experiment-level K-fold cross-validation, while an entropy-production penalty was included in the loss function to discourage violations of the second law of thermodynamics. Across the investigated cross-validation configurations, the global normalized root-mean-square error remained under 3%, with sloshing representing the most demanding regime. A final model trained on the complete database reconstructed the experiments with a global error of 1.6%. These results demonstrate that the proposed framework can extract physically interpretable and computationally efficient closure laws from a limited but information-rich experimental database.
