Invention Description
Wireless communication and sensor systems increasingly use direct digitization of radio-frequency (RF) and intermediate-frequency signals to simplify receiver architectures and reduce analog complexity. Time-interleaved analog-to-digital (ADC) architectures needed for direct RF digitization can introduce mismatches among sub-ADCs, producing in-band spurious tones that degrade signal fidelity. Conventional calibration methods require detailed knowledge of specific error sources, which can make them complex and difficult to adapt across conditions. Prior machine-learning-based ADC calibration approaches may also require re-training for each new input signal, limiting scalability and real-time usability.
Researchers at Arizona State University have developed a machine learning-based digital calibration framework for analog-to-digital converters (ADCs) used in communication and sensing systems. The framework addresses mismatch-induced spurious tones in time-interleaved ADCs by learning error patterns, without requiring prior knowledge of specific error sources. Following one-time training, the model is designed to calibrate varying input signals without additional re-training. Two high-level architectures support different implementation priorities: a network for modeling local and longer-range temporal patterns, and a different network for low-power, real-time calibration. Experimental evaluations reported improvements in signal-to-noise-and-distortion ratio and spurious-free dynamic range, while hardware-aware optimization supports energy-efficient operation.
This deep neural network approach enables blind, one-time training calibration of ADCs without the need for re-training or explicit error modeling.
Potential Applications
- Direct radio frequency and intermediate-frequency digitization in communications systems
- Receiver digital backend calibration for wireless sensing and communication devices
- Low-power and energy-efficient ADC calibration in mobile and IoT hardware
- High-performance instrumentation requiring precise analog-to-digital conversion
- Next-generation RF front-end systems in aerospace, defense, and telecommunications
Benefits and Advantages
- Blind, one-time training without re-training for different input signals
- Error-only prediction enables more efficient calibration
- Simultaneous correction of multiple mismatch effects without explicit error models
- Choice between high-accuracy temporal modeling and lightweight low-power model
- Improved SNDR ratio and SFDR
- Supports reduced numerical precision and model sparsification for energy-efficient hardware deployment
- Eliminates continuous tuning complexity and manual intervention
For more information about this opportunity, please see