VLSI26: NeuroSAR: Physics-informed metastability-aware SAR ADC

[ayan-biswas0412]

Add NeuroSAR: physics-informed metastability-aware SAR ADC repo for SSCS VLSI 2026

This PR introduces NeuroSAR, a physics-informed, metastability-aware SAR ADC waveform engine built as a notebook-centric, open-source project targeting the IEEE SSCS Code-a-Chip / VLSI 2026 program.

NeuroSAR combines a Sky130-inspired SAR ADC abstraction, a PINN surrogate for transient dynamics, and interactive Jupyter visualizations to enable real-time educational exploration of DAC switching, comparator regeneration, and figure-of-merit (FoM) tradeoffs.

---

What’s included

Repository structure

• Add src/, notebooks/, data/, docs/, scripts/, tests/, and assets/ directories for a submission-ready layout.
• Add README, license, requirements.txt, and basic project configuration files.

Physics-informed modeling

• Implement a PyTorch-based PINN surrogate that predicts SAR DAC and comparator waveforms over time and bit index.
• Encode circuit-inspired physics via:
  • KCL-style residuals on DAC-related nodes.
  • Charge-conservation residuals during capacitor switching events.
  • Comparator large-signal regeneration ODE residuals to capture metastability and resolution dynamics.
• Expose configuration hooks for:
  • number of bits, Vref, unit capacitor size, load capacitance,
  • comparator gain, regeneration time constant, and offset,
  • temperature and sampling rate.

Data generation

• Add a lightweight synthetic transient generator for SAR bit-cycling and comparator behavior, so the project runs out-of-the-box.
• Define a clean interface for integrating real Sky130/ngspice transient data (e.g., CDAC + comparator decks) without changing the training or visualization code.
• Store waveforms and design parameters under data/ in a structured tabular format.

Training and evaluation

• Add a full PINN training pipeline:
  • dataset loading and preprocessing,
  • combined data + physics residual loss,
  • configurable loss weights via a central config,
  • checkpoint saving and basic logging.
• Provide evaluation utilities to:
  • compare PINN predictions to reference waveforms,
  • inspect residuals and basic error metrics.

FoM analysis and inverse design

• Implement FoM utilities for:
  • energy per conversion,
  • ENOB proxy from residue/error,
  • metastability dwell time or risk metric,
  • Walden-like SAR FoM.
• Add simple gradient-based inverse-design scaffolding to:
  • adjust design parameters (e.g., unit cap, comparator gain),
  • move towards reduced metastability under energy constraints,
  • visualize optimization trajectories.

Interactive notebooks

• Add Jupyter notebooks to support:
  • dataset generation,
  • PINN training and inspection,
  • waveform visualization,
  • FoM surface exploration,
  • small inverse-design demonstrations,
  • a flagship interactive SAR ADC “lab” with sliders and live plots.
• Use ipywidgets + plotly to:
  • sweep Vin around code boundaries,
  • vary comparator/array sizing,
  • immediately redraw DAC and comparator waveforms and FoMs.

Docs and tests

• Document:
  • the physics model (KCL, charge conservation, comparator ODE),
  • the training plan (splits, losses, checkpoints),
  • a short technical report and submission checklist for SSCS / VLSI.
• Add basic tests to validate:
  • model I/O shapes