OpenAnalogNN — Differentiable Analog Neural Network Simulation from PyTorch to SPICE
77.58% analog accuracy · 28× lower chip variance · 42/42 SPICE match · R²=0.9385 scaling law

---

⚡ One-liner

pip install open-analog-nn

# Then: train analog-aware → map to circuits → simulate SPICE → calibrate → publish

🏆 Why OpenAnalogNN?

The analog ML gap: Software-trained weights must be realized as physical conductances. Manufacturing tolerances, temperature drift, and finite precision destroy accuracy. Prior tools either simulate non-idealities (no SPICE) or generate circuits (no training).

OpenAnalogNN bridges both worlds — the first framework that is simultaneously:

✅ Differentiable	✅ SPICE-Validated	✅ Calibration-Ready	✅ Production
Train through noise, mismatch, offset	42/42 outputs match ngspice at 10⁻⁴	6 methods: affine (best), Bayesian GP, HMAC	PyPI package, 37 tests, arXiv paper

📊 Key Results

SOTA on Fashion-MNIST:

Method	Analog Acc	Chip Std	vs Nature Comms 2026
Standard Deploy	70.66%	2.77%	—
Nature Comms 2026	12.14%	1.36%	Baseline (collapses)
DifferentiableAnalogMLP	77.58%	0.09%	+2.91%, 28× lower variance
+ Affine Calibration	77.87%	0.16%	Best accuracy

SPICE Validation:

  ngspice vs FallbackSolver:  42/42 outputs match at 1e-4  Max diff: 8.87e-5  Mean diff: 3.25e-5

Calibration:

  Affine (OLS):      77.87% accuracy  (best for classification)
  Bayesian GP:       62.0% RMSE reduction  (best for regression, with uncertainty)
  HMAC:              58.7% RMSE reduction  (BLUE-optimal under circuit physics)

Scaling Law (504 runs, R²=0.9385):

  accuracy_drop = 0.130 x D^0.26 x W^0.18 x N^0.86 x exp(-0.35 x log D x log N)

Energy-Accuracy Pareto:

  D=1, W=32 at 7nm:  8980 acc/uJ  (all architectures under 10 uJ per inference)

_{Left: Op-amp differential summing amplifier  ·  Center: Analog scaling law (R²=0.9385)  ·  Right: Energy-accuracy Pareto (28nm–7nm)}

_{Left: Six calibration methods compared  ·  Right: SPICE vs solver parity (42/42 at 10⁻⁴)}

🚀 Quick Start

import torch
from analog_layers import AnalogLinear
from calibration import AffineCalibrator
from circuit_ir.mapping import map_layer_to_circuit
from spice.fallback_solver import FallbackNodalSolver
from validation.metrics import compute_metrics

# 1. Create a differentiable analog layer
layer = AnalogLinear(64, 10, config={
    'noise_sigma': 0.03,           # Weight noise
    'resistor_mismatch': 0.01,      # Pelgrom mismatch
    'saturation_vmax': 2.5,         # Supply rail
    'opamp_offset': 0.002,          # Input offset
    'quantization_bits': 8,         # ADC/DAC resolution
})

# 2. Forward pass (differentiable! gradients flow through non-idealities)
x = torch.randn(32, 64)
y_ideal = torch.randn(32, 10)
y_analog = layer(x)                  # Includes all non-idealities
loss = torch.nn.functional.mse_loss(y_analog, y_ideal)
loss.backward()                      # Works because all ops are reparameterized

# 3. Map trained weights to physical circuit
circuit = map_layer_to_circuit(layer.weight.detach(), layer.bias.detach(), x[0])

# 4. Simulate using SPICE-level solver (matches ngspice at 1e-4)
y_spice = FallbackNodalSolver.solve_closed_form(
    layer.weight.detach(), layer.bias.detach(), x, {
        'resistor_mismatch': 0.01, 'opamp_offset': 0.002,
        'saturation_vmax': 2.5, 'enable_mismatch': True,
        'enable_offset': True, 'enable_saturation': True
    }
)

# 5. Calibrate out systematic errors
cal = AffineCalibrator()
cal.fit(y_spice, y_analog)
y_calibrated = cal.calibrate(y_spice)

# 6. Validate
metrics = compute_metrics(y_analog, y_spice, y_calibrated, labels)
print(f"RMSE reduction: {metrics['rmse_reduction_pct']:.1f}%")

🧩 Features

Non-Idealities (7 types)

Type	Class	Effect
Weight noise	noise_models.py	Temporal Gaussian fluctuation
Resistor mismatch	mismatch.py	Pelgrom σ_R up to 5%
Op-amp offset	analog_linear.py	Amplified by noise gain (1+∑|w|)
Quantization	quantization.py	4–16 bit ADC/DAC
Saturation	saturation.py	Supply rail clamping ±V_max
TCR drift	temperature_dependence.py	Polysilicon: 4.80% over 60°C
Thermal noise	thermal_noise.py	Johnson-Nyquist: ~0.04% at 300K

Calibrators (6 methods)

Method	RMSE ↓	Accuracy	Best For
Affine (OLS)	32.2%	77.87%	Classification
Polynomial	—	—	Mild non-linearity
Bayesian GP	62.0%	73.87%	Regression + uncertainty
HMAC	58.7%	77.58%	Physics-aware correction
Learned MLP	56.9%	—	Non-linear patterns
Ensemble	53.7%	75.40%	Balanced trade-off

Datasets (7)

XOR, Iris, MNIST, Fashion-MNIST, CIFAR-10, SVHN, California Housing

SPICE Infrastructure

• SPICE Autograd — torch.autograd.Function wrapping ngspice with analytical backward
• Fallback Solver — algebraic KCL solution, 1000× faster, matches ngspice at 10⁻⁴
• Exporters — ngspice + LTspice netlist generation
• Multi-layer validation — 2-layer (42/42) and 3-layer cascade (78/102, first documented intermediate saturation)

Novel Contributions

• Hardware Variation Dataset (100-chip Pelgrom population)
• Scaling-law-guided Neural Architecture Search
• Physics-based Energy Model (28nm/14nm/7nm)
• Temperature-aware training (4.80% TCR robustness)
• SPICE Autograd for end-to-end differentiable circuit optimization

🎮 Interactive Demo

pip install open-analog-nn streamlit
streamlit run app.py

Explore 7 datasets, 6 architectures, all calibrators in real-time. Deploy config at app_deploy/ for HuggingFace Spaces.

📦 Install

pip install open-analog-nn

Requires Python 3.10+ and PyTorch 2.0+.

📁 Structure

analog_layers/      # Differentiable non-ideality layers (7 types)
circuit_ir/         # Circuit IR → SPICE exporters (ngspice, LTspice)
spice/              # Runner, autograd, fallback solver, waveform parser
calibration/        # 6 methods: affine, polynomial, Bayesian GP, HMAC, learned, ensemble
validation/         # Metrics, parity plots, residual analysis, statistical tests
datasets/           # 7 dataset loaders with procedural generators
experiments/        # Training pipelines, sweeps, config loader
training/           # Advanced: adversarial, temperature-aware, distributional
nas/                # Analog-aware Neural Architecture Search
energy/             # Physics-based energy efficiency model (28nm–7nm)
huggingface/        # HuggingFace transformer integration (GPT-2, BERT)
benchmarks/         # Full benchmark suite
reports/            # arXiv paper + auto-generated figures
tests/              # 37 tests (37 pass, 2 skipped for HF network access)

📄 Paper

Full paper at reports/arxiv_submission/main.tex with 8 figures, 8 tables, 13 sections.

@software{opennnalog2026,
  title  = {OpenAnalogNN: Differentiable Analog Neural Network Simulation,
            Calibration, and SPICE Validation},
  author = {Varshini C B and OpenAnalogNN Research Group},
  year   = {2026},
  url    = {https://github.com/varshinicb1/AnalogNN},
  doi    = {10.5281/zenodo.XXXXXXX}
}

🤝 Contributing

PRs welcome! See Discussions for roadmap and ideas.

---

⭐ Star this repo — it helps researchers discover analog ML tools!
_{Built with ❤️ for the neuromorphic computing community}