An approach for thermal conductivity measurements in thin films: Combining localized surface topography, thermal analysis, and machine learning techniques

TL;DR

This paper introduces a novel, integrated methodology combining localized surface measurements, substrate correction, and machine learning to accurately determine the thermal conductivity of thin films, improving reliability and scalability.

Contribution

It develops a comprehensive, adaptive approach that combines surface topography, thermal analysis, and machine learning to enhance measurement accuracy of thin film thermal conductivity.

Findings

High predictive accuracy with R^2 of 0.97886 during training

Effective correction for substrate and thickness effects

Scalable framework for diverse materials

Abstract

This study presents a comprehensive methodology for determining the thermal conductivity (TC) of materials with high reliability. The methodology addresses issues such as surface topographical variations and substrate interference by combining Scanning Thermal Microscopy (SThM) with machine learning (ML) models and normalization techniques. Micro- and nanostructural variations in thin films exacerbate measurement inconsistencies, reducing repeatability and reliability. These interconnected challenges highlight the need for a novel, flexible, and adaptive methodology that can comprehensively address the complexities of thin film characterization while maintaining accuracy and efficiency. In this approach, sample surface was divided into fine spatial grids for localized thermal and topographical measurements. A substrate-thickness factor (C factor) was introduced to account for thickness and substrate effects on thin film TC, and high-performance Random Forest regression was used to predict TC across a broad range of materials. The models were trained on a dataset of 2,352 measurements that covered a wide range of material properties and then validated with an additional 980 measurements. They achieved high predictive accuracy, with a $R^2$ of 0.97886 during training and 0.96630 during testing. This approach addresses instrumental limitations and integrates experimental techniques with computational modeling, providing a scalable framework for a wide range of material science applications.