Square cavity coupled heat transfer dataset

Physically Informed Neural Networks (PINNs) integrate physical knowledge into deep learning by mathematically encoding a system of partial differential equations, with successful applications in fluid mechanics and heat transfer. For instance, in the aerospace industry, PINNs are employed to predict the flow and temperature fields within a turbine blade. In this manuscript, a dual cavity model is employed as a simplified representation of the turbine blade, and the coupled heat transfer process within this model is numerically simulated to validate the efficacy of PINNs in this field. However, the strong coupling between solid thermal conductivity and fluid heat transfer renders it challenging for the generalized PINNs to effectively address the coupled heat transfer problem in this field. Consequently, in this manuscript, a PINNs based on a partitioned coupling strategy is employed to numerically simulate the coupled heat transfer problem. The strategy effectively couples the temperature fields in each subregion by dividing the complex coupled problem into multiple subregions and performing numerical simulations in each subregion using PINNs. The coupled heat transfer process is simulated in both cavities using COMSOL® software, and the predicted data set is exported for use in the model. The size of the dataset is 200×420×1000 (x×y×t), and the dataset is divided into a training set (0~80,000,000) and a validation set (80,000,000~84,000,000).

物理信息神经网络（Physically Informed Neural Networks, PINNs）通过将偏微分方程组以数学形式编码，将物理知识融入深度学习，目前已在流体力学与传热领域取得成功应用。例如，在航空航天工业中，PINNs被用于预测涡轮叶片内部的流场与温度场。本文采用双腔模型作为涡轮叶片的简化表征，并对该模型内的耦合传热过程进行数值模拟，以验证PINNs在该领域的应用有效性。然而，固体导热与流体换热之间的强耦合特性，使得广义PINNs难以有效求解该领域的耦合传热问题。为此，本文采用基于分区耦合策略的PINNs，对该耦合传热问题开展数值模拟。该策略通过将复杂耦合问题拆解为多个子区域，并利用PINNs在各子区域内开展数值模拟，实现了各子区域温度场的有效耦合。本文采用COMSOL®软件对两个腔体内的耦合传热过程进行模拟，并导出预测数据集用于模型训练。该数据集的尺寸为200×420×1000（x×y×t），并被划分为训练集（0~80,000,000）与验证集（80,000,000~84,000,000）。