Page 26 - december22_112922.indd
P. 26
TAP TIMES -26- DECEMBER 2022
Terahertz camera (from 25) Argonne Lab researchers
study why materials corrode
regular camera,” he says. In the team’s
experiments, the device was able to detect LEMONT, IL. — Researchers at the U.S.
terahertz pulses at low intensity levels that Department of Energy’s (DOE) Ar-
surpassed the capability of today’s large gonne National Laboratory have de-
and expensive systems. veloped a new technique that acceler-
ates the solving of material structures
Detector capabilities from patterns uncovered in X-ray
The researchers demonstrated the ca- experiments.
pabilities of the detector by taking tera-
hertz-illuminated pictures of some of the Battery charging
structures used in their devices, such as The technique allows researchers to study
the nano-spaced gold lines and the ring- certain properties, such as corrosion or
shaped slits used for the polarized detec- battery charging and discharging in real
tor, proving the sensitivity and resolution time.
of the system.
The technique, called AutoPhaseNN, is
Developing a practical terahertz camera based on a method called machine learn-
requires a component that produces tera- ing, which trains an algorithm on certain
hertz waves to illuminate a subject, and experimental data and then uses it to
another that detects them. choose the most likely outcome of the cur-
rent experiment.
On the latter point, current terahertz de-
tectors are either very slow, because they The data used in this case are created by
rely on detecting heat generated by the shining ultrabright X-ray beams from Ar-
waves striking a material, and heat propa- gonne’s Advanced Photon Source (APS) on
gates slowly, or they use photodetectors a material and capturing the light as they
that are relatively fast, but have very low bounce off, a process called diffraction.
sensitivity. The APS is a DOE Office of Science user
facility at Argonne.
Cost of detectors
In addition, until now, most approaches Machine learning
have required a whole array of terahertz AutoPhaseNN is an example of an “unsu-
detectors, each producing one pixel of pervised” machine learning, which means
the image. “Each one is quite expensive,” that the computer algorithm learns from
Shi says, so “once they start to make a its own experience how to do a computa-
camera, the cost of the detectors starts to tion more accurately and efficiently, with-
scale up really, really quickly.”
out having to be trained with labeled solu-
tions that have already been figured out,
While the researchers say they have a process which usually involves human
cracked the terahertz pulse detection intervention.
problem with their new work, the lack anl.gov
of good sources remains — and is be-
ing worked on by many research groups
around the world. TAP TIMES is looking for articles that enlighten and
mit.edu educate our readers. E-mail roniscoff@gmail.com
