Abstract

The traditional model-driven methods are not much efficient to predict the viscosity of nanofluids accurately. This study presents a novel approach of using physics-guided deep learning technique for predicting viscosity of water-based nanofluids from large dataset containing both experimental and simulated data of spherical oxide nanoparticles $\rm{Al2O3}$, $\rm{CuO}$, $\rm{SiO2}$, and $\rm{TiO2}$. Further, this study introduces a novel methodology of combining deep learning methods and physics-based models to leverage their complementary strengths. To the best of the author’s knowledge, theory-guided deep learning prediction model was never used to predict viscosity before. The theory-guided deep neural networks (TGDNN) model is trained by minimizing the mean square error (MSE) and regularization terms using Adam optimization technique. The investigations reveal that the values of R², RMSE, and AARD% are, respectively, 0.999868, 0.001143, and 2.198887 on experimental testing dataset. The TGDNN model learns non-linear relationship among the input variables from the training data. Additionally, the results show that the proposed method performed better than the other well-known existing theoretical and computer-aided models to predict the viscosity in wide range with high level of accuracy.

中文翻译：

---

基于物理的智能模型，用于通过深度学习预测包含纳米颗粒的纳米流体的粘度

摘要

传统的模型驱动方法不能有效地准确预测纳米流体的粘度。这项研究提出了一种使用物理指导的深度学习技术从包含球形氧化物纳米颗粒$ \ rm {Al2O3} $，$ \ rm {CuO} $的实验数据和模拟数据的大型数据集中预测水基纳米流体粘度的新方法，$ \ rm {SiO2} $和$ \ rm {TiO2} $。此外，本研究介绍了一种结合深度学习方法和基于物理的模型以利用其互补优势的新颖方法。据作者所知，理论指导的深度学习预测模型以前从未用于预测粘度。通过使用Adam优化技术最小化均方误差（MSE）和正则项来训练理论指导的深度神经网络（TGDNN）模型。在实验测试数据集上，R ²，RMSE和AARD％分别为0.999868、0.001143和2.198887。TGDNN模型从训练数据中学习输入变量之间的非线性关系。另外，结果表明，所提出的方法比其他已知的现有理论模型和计算机辅助模型表现更好，能够以较高的准确度预测大范围的粘度。