Background
Noise-shaping SAR ADCs achieve Σ-Δ comparable precision at low oversampling rates, but their closed-loop structure is extremely sensitive to capacitor mismatch, integrator coefficients, and comparator offset. Conventional LMS background calibration operates linearly in coefficient space, making it difficult to simultaneously correct nonlinear distortion components from different orders that are mutually coupled.
Approach
Treat the ADC as an unknown nonlinear mapping $f: \mathbf{d} \to v_\text{in}$ and learn its inverse mapping using KAN’s (Kolmogorov–Arnold Network) differentiable spline nodes. Compared to MLPs, KAN shifts activation functions to edges with nodes performing only addition, making each nonlinearity visualizable and prunable for physical diagnostics.
Current Progress
- Complete Python behavioral-level model (sampling, quantization, noise-shaping NTF, mismatch injection)
- KAN calibrator training pipeline + benchmark: improving from 9.4 → 12.1 ENOB under ±2% capacitor mismatch
- Three-way comparison with traditional polynomial and MLP calibrators
Next Steps
- Quantize the KAN calibration mapping into lookup tables and evaluate hardware cost of RTL implementation
- Co-simulation: replace the behavioral model with actual SAR backend designed in TSMC55, validating transistor-level consistency