基于小样本学习的水基纳米流体导热系数预测

Prediction of thermal conductivity of nanofluids poses a notable challenge because of the inherent nonlinear characteristics of thermal conductivity. Herein, we introduce a small-sample observation series prediction model as a gray model and employ a “DAN with three branches for regression” (DAN-TR) network to predict the thermal conductivity of nanofluids using the “migration learning + numerical fitting pseudolabeling” method. Given the limited sample size, direct fitting would result in underfitting and decreased network accuracy. Therefore, a substantial amount of pseudolabeled data is initially generated by fitting the actual data with the gray model. The model is then trained on these pseudolabeled data and subsequently fine-tuned with actual data, thereby obtaining more accurate results. After optimizing the DAN-TR small-sample learning network, the proposed method delivers good prediction performance both for the CuO-H₂O and Al₂O₃-H₂O systems.