Utilizing machine learning in the three-omega method to predict thermophysical properties with low variation

Abstract

The three-omega method offers a precise and simple technique for measuring the thermal conductivities of a wide range of materials. Utilizing nonlinear data enhances the determination of multiple thermophysical properties at the micro- and nanoscales. However, challenges such as dependence on initial guesses and convergence to local minima render nonlinear fitting processes unstable. This study introduces a machine learning-based approach that addresses these limitations by pre-learning a globally optimal solution. We developed a methodology that processes three-omega signals using machine learning to determine various thermophysical properties. The approach involved compiling results from the analytical three-omega model and augmenting them with Gaussian noise to create a robust machine learning model. Subsequently, a neural network was trained on this dataset. Testing confirmed the model's accurate predictions of volumetric heat capacity and thermal conductivity. Comparative analyses between machine learning-based and conventional fitting methods on the sample data extracted from the test set demonstrated the superior stability of the machine learning model in predicting thermophysical properties. Furthermore, experimental validation was carried out by applying our machine learning model to real-world data from glass and liquid samples. The correspondence between the predicted values and the theoretical curve of the third harmonic voltage affirmed the model's applicability and accuracy in practical scenarios. These findings validate the potential of our machine learning model for robustly predicting thermophysical properties based on actual measured data.