【摘要】 通过少量离散的测点准确重构完整温度场可以方便快捷地大幅扩充测量数据，在航空航天等领域具有重要意义。本文以三维平板气膜冷却温度场重构问题为例，提出了一种基于迁移学习和物理约束(TL-PINN)的三维温度场重构方法。首先将导热方程加入神经网络损失函数作为约束，并通过其他工况的丰富数据进行预训练，再迁移至目标域内，通过少量测点微调训练得到最终模型。基于数值仿真数据进行模型的训练和测试，结果显示，相比于常规前馈神经网络和物理约束神经网络，TL-PINN具有更低的温度场预测误差和更少的模型训练时间。并且考察了不同可训练层和源工况条件对最终模型温度预测能力的影响，发现仅训练部分隐藏层可节约训练时间，源工况条件对模型的影响较小。更多还原

【Abstract】 Reconstructing the temperature field accurately based on a small number of discrete measurements can efficiently expand the measurement database, which is of great significance in aerospace and many other fields. In present study, a transfer learning and physics-informed neural network(TL-PINN) is proposed to reconstruct the temperature field of a three-dimensional flat plate with film cooling. Firstly, the heat conduction equation is added to the neural network’s loss function as a constraint, and the network is pre-trained by sufficient data from other working conditions.Then, the network is transferred to the target working condition, and the final model is obtained by fine-tuning with a small amount of measurement data. The model is trained and tested with the numerical simulation data. The results show that TL-PINN has lower temperature prediction error and less training time compared with the conventional back propagate neural network and physicsinformed neural network. The effects of different trainable layers and source working conditions on the prediction ability of the final model are also investigated. It is found that only training some hidden layers can save training time, and the source working conditions have less influence on the model.更多还原