Machine Learning THz Photonics - Tera Learn

Terahertz (THz) radiation provides advantages compared to other spectral regions for applications in the field of spectroscopy, security and imaging. However, the generation and detection of THz radiation is a technological challenge. Quantum cascade lasers (QCLs), semiconductor devices whose optical transitions are defined by the quantized energy levels in the nanostructures, are a promising THz source. Thus, the emission wavelength can be designed in a wide range. Large progress has been made in recent years regarding the maximum operating temperature and the development of ultrabroadband active regions. In this project we will take advantage of these novel ultrabroadband structures. To unlock the potential of ultrabroadband gain we will work with new cavity structures which do not predefine the emission wavelength a priori. To control lasing action in these structures, we will apply Artificial Neural Networks (ANN), as they can solve the complex problems light matter interaction rapidly. The combination of ANNs and THz QCLs allows elegant solutions for various problems of THz QCL applications. ANNs will be used for the creation of a fast beam steering mechanism, the generation of arbitrarily shaped THz spectra for spectroscopy and spectrally resolved object recognition in the THz region. We envision a hardware implementation of the ANNs for hyperspectral recognition, as this avoids high computational effort. Our methods for beam steering, spectroscopy and object recognition are thus easily transferable to applications. We will develop highly tunable random lasers based on dual-stack InGaAs/InAlAs active regions. We will also develop cavity designs for quantum cascade random lasers, which allow a high degree of external modulation. We will implement tunable weights by all-optical modulation of the THz beam. For beam steering, we will use a random laser cavity, where different lasing modes emit in different directions. The ANN maps the modulation of the THz laser to its emission direction. To generate arbitrary laser spectra, the ANN fits the mode interaction dynamics to shape the modes toward the desired frequencies. For object recognition and material determination, the ANN is trained to recognize shapes. We will use a single-pixel based approach in combination with a highly tunable laser source to recognize trained objects as well as spectral training iterations to determine its material composition. Typical cavity concepts rely on Fabry-Pert resonances or Whispering-Gallery modes, which limits the output spectra to regular spacings. We motivate a resonator design which does not rely on cavity modes but on random modes, which are shaped by NIR modulation. This highly tunable cavity concept enables the application of artificial neural networks to shape the emission properties quickly.