Predicting Interfacial Thermal Resistance by Ensemble Learning

Mingguang Chen, Junzhu Li, Bo Tian, Yas Mohammed Al-Hadeethi, Bassim Arkook, Xiaojuan Tian, Xixiang ZhangComput 2021 · DOI

TL;DR

This paper tackles the challenge of accurately predicting interfacial thermal resistance (ITR), a crucial property for thermal management in various applications. By employing an ensemble learning approach that combines XGBoost, Kernel Ridge Regression, and Deep Neural Networks, the authors achieve robust predictions and identify promising high-melting point, high-ITR material systems for future research and development.

What Problem Does This Paper Address?

Interfacial thermal resistance (ITR) is a critical factor influencing the thermal properties of materials, impacting everything from integrated circuits to engine turbines. Accurately predicting ITR is vital for designing advanced materials and ensuring effective thermal management. However, ITR is influenced by a multitude of factors (dozens, in fact), making traditional models struggle with high-dimensional prediction. Existing models like the Acoustic Mismatch Model (AMM) and Diffuse Mismatch Model (DMM) have simplified assumptions that limit their applicability, while molecular dynamics simulations, though more comprehensive, are computationally expensive. This complexity and the high dimensionality of the problem necessitate more advanced predictive tools.

Methods

The authors propose an ensemble learning framework to overcome the limitations of individual models and the challenge of overfitting with limited datasets. The process involves:

  • Exploratory Data Analysis (EDA) and Data Visualization: To gain a comprehensive understanding of the raw data and identify relationships between descriptors and ITR.
  • Descriptor Selection: Utilizing XGBoost's feature importance scores and Pearson correlation coefficients to identify the most relevant descriptors, reducing dimensionality and improving model conciseness. They selected 17 key descriptors out of an initial 35.
  • Individual Model Training: Training three distinct models:
  • XGBoost: A gradient boosting algorithm known for its speed and accuracy.
  • Kernel Ridge Regression (KRR): A well-established machine learning method that combines ridge regression with the "kernel trick" for nonlinear relationships.
  • Deep Neural Network (DNN): A multi-layered artificial neural network designed to capture complex, non-linear patterns.
  • Ensemble Model Creation: Combining the predictions of the XGBoost, KRR, and DNN models through averaging. This ensemble approach aims to reduce variance, improve accuracy, and enhance robustness.

Datasets

The study utilized two datasets:

  • "training dataset for ITR prediction.xlsx": This dataset, containing 692 instances, was collected from the literature and organized by the Xu group.
  • "descriptor dataset.xlsx": This dataset was used to construct material systems.

The authors note that the dataset is relatively small, which is a common challenge in this field and a primary motivation for using ensemble learning to combat overfitting.

Key Insights and Ideas

  • Ensemble Learning for Robustness: The core idea is that combining multiple diverse models (XGBoost, KRR, DNN) through averaging can lead to more robust and accurate predictions than any single model, especially when dealing with limited data and potential overfitting.
  • Descriptor Importance: The study successfully identified key descriptors that significantly influence ITR, such as binding energy, melting point, density, and heat capacity. This insight is valuable for both predictive modeling and understanding the underlying physics of thermal transport.
  • Dimensionality Reduction is Key: By carefully selecting a subset of descriptors (17 out of 35), the authors were able to build concise models that maintained high predictive performance, demonstrating that not all features are equally important and that feature selection can improve efficiency and interpretability.
  • Predicting Novel Material Systems: The developed ensemble model was used to predict novel high-melting point, high-ITR material systems, offering a valuable resource for experimentalists and engineers.

Key Results

  • The ensemble model achieved a high R² score of 0.87 and a low RMSE of 9.34 on the test dataset, outperforming individual models.
  • The predicted ITR for the Pb/Diamond material system showed strong agreement with experimentally reported values, validating the model's accuracy.
  • The study identified the top 20 high melting point, high ITR material systems, with many composed of carbon-based substrates like diamond, graphite, and graphene.

Limitations and Future Work

  • Dataset Size: The authors acknowledge that the training dataset is small and cannot encompass all possible material systems. They suggest that combining more data from experiments is crucial.
  • Limited 2D Materials: The current dataset contains only a small fraction (2.9%) of two-dimensional material systems, which they aim to address in future work.
  • Future Directions: The researchers plan to:
  • Develop a comprehensive database for ITR data.
  • Collect additional ITR data from recent literature and experiments.
  • Investigate multicomponent (>2) high melting point, high ITR material systems, potentially through synthesis of corresponding nanomaterial compounds.

Summary updated

BibTeX
@article{chen2021predicting,
  author = {Mingguang Chen and Junzhu Li and Bo Tian and Yas Mohammed Al-Hadeethi and Bassim Arkook and Xiaojuan Tian and Xixiang Zhang},
  title = {{Predicting Interfacial Thermal Resistance by Ensemble Learning}},
  journal = {comput.},
  year = {2021},
  doi = {10.3390/computation9080087},
}
Figure