Date published: June 4, 2026
Summary: A spectral kernel machine uses a tunable photodetector and spectral encoder to perform machine learning directly during light detection, allowing for real-time, low-power identification of materials and objects without external data processing.
Applications:
- Semiconductor wafer thickness metrology
- Thin film coating metrology
- Optical sorting/factory automation
- UAV/satellite based remote sensing
- Automated target object identification
Advantages/Benefits:
- Low power consumption
- Real-time processing speed
- Eliminates data transport bottlenecks
- High-accuracy analog inference
- In-situ adaptability to field conditions
Background:
Hyperspectral imaging is critical for identifying materials in fields like remote sensing and chemometrics. There is a growing need for real-time, low-power spectral intelligence that can operate efficiently on mobile platforms to instantly identify specific optical signatures.
Traditional systems capture massive 3D data hypercubes for external digital post-processing. This approach creates severe data bottlenecks, demands high-bandwidth transport, and results in excessive power consumption and low frame rates, hindering rapid on-device inference.
Technology Overview:
Scientists at Berkeley Lab and UCLA have developed a spectral kernel machine integrating a spectral encoder and an electrically tunable photodetector to perform analog machine learning inference during photodetection. Using materials like silicon or black phosphorus-MoS2, it co-modulates incident light to execute mathematical operations. The device outputs a photocurrent directly representing analysis results and uses genetic algorithms for in-situ training.
This technology is differentiated by shifting spectral imaging from “capture-then-process” to “detect-as-inference.” Unlike traditional systems capturing dense 3D hypercubes for external processing, it physically encodes machine learning kernels into the photodetector’s response. This eliminates data bottlenecks and high-bandwidth transport, enabling real-time, low-power operations for mobile platforms. The device consumes more than 100 times less power, operates more than 100 times faster, and achieves more than 90% classification accuracy for multiple practical multispectral/hyperspectral machine learning tasks.
The scientists have successfully demonstrated visible to mid-infrared (MIR) spectral inference at the single-pixel level. While mid-infrared scalability is currently under exploration, development is also underway to transition this technology into high-resolution megapixel imaging arrays.
Development Stage: TRL3
Inventors:
Aydogan Ozcan, UCLA
Yuhang Li, UCLA
Status: Patent pending
Opportunities: Available for licensing and / or collaborative research
For More Information:
Zhang, D. et al (2025). Spectral kernel machines with electrically tunable photodetectors, Science, Vol 390, Issue 6776. DOI: 10.1126/science.ady6571
