| Περίληψη |
This thesis explores new, modern approaches to material parameter extraction in Terahertz Time-
Domain Spectroscopy (THz-TDS) by utilizing machine learning techniques. Traditionally, the anal-
ysis of THz-TDS data relies on computationally intensive methods such as the Newton-Raphson
algorithm to calculate the complex refractive index of materials. We propose a data-driven ap-
proach utilizing neural networks to predict the real and imaginary parts of the refractive index. Two
neural network models were developed and compared: a Single Model that predicts both compo-
nents simultaneously, and a Double Model that predicts them separately. The models were trained
using synthetic datasets generated through a signal generation algorithm that simulates data taken
from THz-TDS experiments. The best-performing models reached mean absolute percentage errors
(MAPE) of approximately 2.17% and 1.55% for the real part of the refractive index and 15.26%
and 17.71% for the imaginary part, respectively. While the real part predictions were deemed sat-
isfactory, the imaginary part predictions had higher errors because of the limitations in the dataset
generation and the inherent challenges of predicting smaller values.
Additionally, real data predictions were compared against traditional methods, revealing fur-
ther insights. Although the neural network models were able to follow general trends, significant
discrepancies emerged for the real part of the refractive index. The imaginary part predictions, how-
ever, aligned more closely with the traditional values, although some oscillations and overshooting
were still evident. These findings suggest that further refinement is required, especially for the real
part predictions, to improve the overall accuracy of the models.
This work demonstrates the potential of machine learning applications to enhance the efficiency
and accuracy of THz-TDS data analysis, paving the way for more advanced applications in material
characterization.
|
THz-TDS material parameter extraction using neural networks
