Presentation Information
[8a-C213-5]Quantitative doping-density estimation from terahertz emission waveforms via physics-informed machine learning
〇Yuta Akase1, Iwao Kawayama1, RAZANOELINA Manjakavahoaka1 (1.Kyoene Univ)
Keywords:
terahertz,machine learning
The doping density of a semiconductor p-n junction is reflected in the terahertz
(THz) emission waveform excited by a femtosecond laser. We propose a
physics-informed machine learning (PINN) framework that inversely estimates the
donor and acceptor densities (N_D, N_A) from THz emission waveforms. Training data
were generated by Monte Carlo device simulation built on ViennaEMC.
A physics-consistency loss derived from the Poisson equation constrains the predicted
densities so that the reconstructed built-in potential and peak field agree with
the data. The model was trained and evaluated solely on simulated waveforms,
augmented with an instrument response, timing jitter, and additive white Gaussian
noise to mimic measurement conditions. The network estimated N_D with R2=0.994 (MAE
0.031 dex) and remained robust over 20-40 dB SNR, while N_A was less accurate
(R2=0.92). These results show the feasibility of non-contact doping-density
estimation from THz emission waveforms.
(THz) emission waveform excited by a femtosecond laser. We propose a
physics-informed machine learning (PINN) framework that inversely estimates the
donor and acceptor densities (N_D, N_A) from THz emission waveforms. Training data
were generated by Monte Carlo device simulation built on ViennaEMC.
A physics-consistency loss derived from the Poisson equation constrains the predicted
densities so that the reconstructed built-in potential and peak field agree with
the data. The model was trained and evaluated solely on simulated waveforms,
augmented with an instrument response, timing jitter, and additive white Gaussian
noise to mimic measurement conditions. The network estimated N_D with R2=0.994 (MAE
0.031 dex) and remained robust over 20-40 dB SNR, while N_A was less accurate
(R2=0.92). These results show the feasibility of non-contact doping-density
estimation from THz emission waveforms.