Machine learning optimises infrared heating for sensors

October 26, 2021 // By Nick Flaherty
Machine learning optimises infrared heating for sensors
Researchers in the US have developed a novel approach to design and fabricate thin-film infrared light sources with an open source machine learning methodology that reduces the optimization time for these devices from weeks or months to a few minutes.

The ability to develop inexpensive, efficient, designer infrared light sources could revolutionize molecular sensing technologies. Additional applications include free-space communications, infrared beacons for search and rescue, molecular sensors for monitoring industrial gases, environmental pollutants and toxins.

A research team at Vanderbilt and Penn State uses simple thin-film deposition, one of the most mature nano-fabrication techniques, aided by key advances in materials and machine learning.

Most thermal emitters with a custom spectral output have required patterned nanostructures fabricated with high-cost, low-throughput methods. Instead, the research team led by Joshua Caldwell, Vanderbilt associate professor of mechanical engineering, and Jon-Paul Maria, professor of materials science and engineering at Penn State, uses cadmium oxide in concert with a one-dimensional photonic crystal fabricated with alternating layers of dielectrics referred to as a distributed Bragg reflector.

The combination of these multiple layers of materials gives rise to a so-called “Tamm-polariton,” where the emission wavelength of the device is dictated by the interactions between these layers. Until now, such designs were limited to a single designed wavelength output. But creating multiple resonances at multiple frequencies with user-controlled wavelength, linewidth, and intensity is imperative for matching the absorption spectra of most molecules.

Material design has been challenging and computationally intense. Because advanced applications require functionality at multiple resonances, the new process had to drastically shorten design time. A typical device, for example, would contain tens to hundreds of designable parameters, creating high customization demands requiring unrealistic computation times. For instance, in a scenario that independently optimizes nine parameters, sampling 10 points per parameter, the simulations would take 15 days assuming 100 simulations each second. Yet, with more parameters, the time increases exponentially—11 and 12 parameters would require three and 31 years, respectively.

To address this challenge, PhD student Mingze He, lead author of the paper, proposed an inverse design algorithm that computes an optimized structure within minutes on a consumer-grade desktop. Further, this code could provide the ability


Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.