[Paper Review] Transfer-learning-based Surrogate Model for Thermal Conductivity of Nanofluids

Saeel S. Pai, Abhijeet Banthiya|arXiv (Cornell University)|Jan 2, 2022

Nanofluid Flow and Heat Transfer4 citations

TL;DR

This paper proposes a transfer learning-based surrogate model using artificial neural networks to predict the thermal conductivity of nanofluids with spherical particles across 32 particle-fluid combinations. By pre-training on lower-accuracy correlation-generated data and fine-tuning with limited high-quality experimental data, the model achieves a goodness of fit of 0.93, significantly outperforming baseline models trained only on experimental data (R² = 0.83).

ABSTRACT

Heat transfer characteristics of nanofluids have been extensively studied since the 1990s. Research investigations show that the suspended nanoparticles significantly alter the suspension's thermal properties. The thermal conductivity of nanofluids is one of the properties that is generally found to be greater than that of the base fluid. This increase in thermal conductivity is found to depend on several parameters. Several theories have been proposed to model the thermal conductivities of nanofluids, but there is no reliable universal theory yet to model the anomalous thermal conductivity of nanofluids. In recent years, supervised data-driven methods have been successfully employed to create surrogate models across various scientific disciplines, especially for modeling difficult-to-understand phenomena. These supervised learning methods allow the models to capture highly non-linear phenomena. In this work, we have taken advantage of existing correlations and used them concurrently with available experimental results to develop more robust surrogate models for predicting the thermal conductivity of nanofluids. Artificial neural networks are trained using the transfer learning approach to predict the thermal conductivity enhancement of nanofluids with spherical particles for 32 different particle-fluid combinations (8 particles materials and 4 fluids). The large amount of lower accuracy data generated from correlations is used to coarse-tune the model parameters, and the limited amount of more trustworthy experimental data is used to fine-tune the model parameters. The transfer learning-based models' results are compared with those from baseline models which are trained only on experimental data using a goodness of fit metric. It is found that the transfer learning models perform better with goodness of fit values of 0.93 as opposed to 0.83 from the baseline models.

Motivation & Objective

• Address the lack of a universal theoretical model for predicting the anomalous thermal conductivity enhancement in nanofluids.
• Overcome the challenge of limited high-quality experimental data for training robust predictive models.
• Develop a data-efficient surrogate model that leverages both low-accuracy correlations and high-accuracy experimental data.
• Improve prediction accuracy for thermal conductivity across diverse particle-fluid combinations using transfer learning.
• Demonstrate the effectiveness of transfer learning in enhancing machine learning models for complex, non-linear thermal phenomena in nanofluids.

Proposed method

• Train artificial neural networks using transfer learning: first on a large dataset of correlation-generated data (lower accuracy), then fine-tuned on a smaller set of experimental data (higher accuracy).
• Use 32 particle-fluid combinations (8 particle materials × 4 base fluids) to evaluate model generalization across diverse nanofluid systems.
• Employ a two-stage training strategy: coarse-tuning on synthetic correlation data to learn broad patterns, followed by fine-tuning on real experimental data to refine predictions.
• Utilize a goodness of fit metric (R²) to quantitatively compare the performance of transfer learning models against baseline models trained exclusively on experimental data.
• Apply standard deep learning architectures with hyperparameter tuning to optimize model convergence and predictive accuracy.
• Ensure model robustness by validating predictions across multiple particle types and base fluids, including water, ethylene glycol, and other common fluids.

Experimental results

Research questions

• RQ1 Can transfer learning improve the predictive accuracy of surrogate models for nanofluid thermal conductivity compared to models trained only on experimental data?
• RQ2 How effective is the use of low-accuracy correlation data as a pre-training source for neural networks in the context of nanofluid thermal conductivity prediction?
• RQ3 To what extent does the transfer learning approach generalize across diverse particle-fluid combinations in nanofluids?
• RQ4 What is the quantitative improvement in model performance when using transfer learning versus traditional supervised learning on limited experimental data?
• RQ5 Can hybrid data training (correlations + experiments) lead to more robust and accurate surrogate models for complex, non-linear thermal phenomena?

Key findings

• The transfer learning-based surrogate model achieved a goodness of fit (R²) of 0.93, significantly outperforming baseline models trained solely on experimental data, which achieved an R² of 0.83.
• The model demonstrated strong generalization across 32 distinct particle-fluid combinations, indicating robustness to variations in material properties.
• Pre-training on correlation-generated data enabled faster convergence and improved parameter initialization, leading to better final performance despite lower data quality.
• Fine-tuning on limited experimental data effectively corrected systematic biases from correlation-based predictions, enhancing accuracy.
• The hybrid training strategy reduced overfitting and improved model stability, especially given the scarcity of high-quality experimental data.
• The results confirm that transfer learning is a viable and effective approach for developing accurate surrogate models in data-scarce scientific domains like nanofluid thermophysics.