Back
Science

Researchers Advance AI-Integrated Spectral Sensing Technologies

Generated on: Last updated: 2 sources
View source

AI-Powered Spectral Sensing: New Innovations from UC Davis and Berkeley Lab

Recent research from two distinct institutions, the University of California Davis (UC Davis) and the Department of Energy's Lawrence Berkeley National Laboratory (Berkeley Lab), has introduced new developments in AI-powered spectral sensing. These innovations aim to integrate artificial intelligence algorithms directly into sensor technology, enabling real-time material identification and analysis, and addressing limitations found in traditional spectral imaging and spectrometry tools.

Background on Spectral Sensing

Traditional spectral sensing tools are often slow, power-intensive, and bulky, requiring large instruments or separate computational modules.

Traditional spectrometers and spectral imaging systems typically require large, expensive laboratory instruments or separate computational modules. These systems often spatially spread light to visualize chemical composition or send vast amounts of data to digital processors, which can result in slow, power-intensive, and bulky setups. The new research focuses on miniaturizing these capabilities and embedding computation directly within the sensing mechanism.

UC Davis Develops Sand-Sized Spectrometer-on-a-Chip

Researchers at UC Davis have developed a spectrometer-on-a-chip, described in Advanced Photonics, that is approximately 0.4 square millimeters in size. This device is designed to reconstruct light spectra using a novel approach that deviates from traditional spatial light spreading.

The UC Davis chip incorporates 16 distinct silicon detectors, each engineered to respond uniquely to incoming light. An artificial intelligence (AI) component is then used to interpret these varied responses and reconstruct the original light spectrum. Key technological components include:

  • Photon-Trapping Surface Textures (PTSTs): These textures were engineered onto standard silicon photodiodes to enhance the absorption of near-infrared (NIR) light. NIR light is relevant for applications such as biomedical imaging due to its tissue penetration capabilities. The PTSTs cause NIR photons to scatter within the thin silicon layer, increasing absorption and extending the chip's spectral sensitivity. The architecture also includes high-speed sensors for ultra-fast photon lifetime measurement.
  • Fully Connected Neural Network: This AI network is trained using thousands of examples to establish relationships between the encoded, noisy signals from the 16 detectors and the pure light spectrum. It can reconstruct the spectrum with an approximate resolution of 8 nanometers, which researchers state removes the necessity for bulky optical components.

The resulting system exhibits high sensitivity and resistance to noise, maintaining signal clarity even in the presence of electrical interference. UC Davis researchers anticipate that this technology could enable integrated, real-time hyperspectral sensing for applications such as advanced medical diagnostics and environmental remote sensing.

The UC Davis spectrometer-on-a-chip, measuring just 0.4 square millimeters, leverages AI and novel photon-trapping textures to achieve high-resolution spectral reconstruction without bulky optics.

Berkeley Lab Introduces AI-Enhanced Sensor for Real-Time Identification

Separately, researchers at the Department of Energy's Lawrence Berkeley National Laboratory (Berkeley Lab) have developed an AI-enhanced sensor designed for real-time identification and characterization of objects and materials. This technology integrates artificial intelligence algorithms directly into the sensor itself, addressing data-processing limitations in current spectral imaging tools.

The Berkeley Lab sensor aims to improve the speed, resolution, and power efficiency of spectral machine vision technologies by performing AI computation and spectral analysis during the image capture process. This approach utilizes photodetection as a physical computational process:

  • Light intensity is mapped to an electrical current.
  • The sensor's responsivity to light can be adjusted to highlight specific spectral signatures.
  • The resulting electrical current directly serves as an inference about the image's spectral content.

This computational process mathematically resembles algorithms used for digital machine learning, allowing the sensor to function as a machine learning computer. The sensor is trained by being exposed to numerous examples of spectral signatures, which enables it to identify specific features in new images without requiring digital post-processing of data.

Experimental demonstrations using black phosphorus photodiodes have shown the sensor's capability in various tasks, including:

  • Identifying oxide layer thicknesses in semiconductor samples.
  • Determining hydration states in plant leaves.
  • Performing object segmentation in optical images.
  • Detecting transparent chemicals.

Developers at Berkeley Lab anticipate broader applications for these smart sensors in advanced optical sensing.

Berkeley Lab's AI-enhanced sensor integrates machine learning directly into the photodetection process, enabling real-time spectral identification and analysis during image capture, significantly improving speed and power efficiency.