05. Micro- and Nano-Scale Transport, MEMS

Increasing heat generation density in semiconductor devices is critical for stable operation in integrated circuits and power devices. It is desirable to obtain internal information via non-contact and non-destructive methods to evaluate the temperature distribution and thermal properties in the depth direction of semiconductors. In addition, the lattice mismatch between substrate and grown materials causes defects in the thickness direction during crystal growth. Therefore, a method for evaluating the thermal properties of thin films with different crystallinities in the thickness direction is also significant. As a non-destructive measurement method, frequency-domain thermoreflectance (FDTR) is one of the powerful tools for measuring the thermal properties of nano/microscale samples. In FDTR measurement, a metal transducer layer is deposited on a sample surface. A sample is periodically heated with a pump laser, and its temperature response on a transducer surface is measured with a probe laser. A phase delay between a pump laser and a probe laser includes sample thermophysical information, thus thermal conductivities can be obtained by fitting measurement data by a theoretical thermal dissipation model. And thermal penetration depth in a sample can be changed by modulating a pump laser frequency. In this study, we propose a machine learning method to analyze the thermal conductivity profile in the depth direction of a sample. We accumulated calculation results from models assuming various thermal conductivity profiles in the depth direction and created training datasets. By training these datasets using a neural network with dropout, we have developed and verified a model capable of estimating the thermal conductivity profile in the depth direction from the FDTR measurements.

Prof.

Toshiyuki Tsuchiya