qtaml-quantum-tunneling-ml

name: qtaml-quantum-tunneling-ml description: "Quantum Tunneling-Aware Machine Learning (QTAML) — derives deployment-time weight-error distribution from WKB approximation, provides closed-form Tunneling-Aware Compensation (TAC) algorithm that reaches 95% clean accuracy with 3.4-33.6x less ECC overhead. Bridges semiconductor physics with noise-aware ML. arXiv:2606.00741" tags: ["machine-learning", "quantum-physics", "hardware-aware", "wkb-approximation", "noise-modeling", "deployment-robustness"] related_skills: ["ai-power-profiling", "quantization-spiking-neural-networks-beyond-accuracy", "quantized-snn-hardware-optimization"] arxiv_id: "2606.00741"

QTAML: Quantum Tunneling-Aware Machine Learning

Paper: "Quantum Tunneling-Aware Machine Learning: Physics-Derived Noise Models for Robust Deployment"

arXiv: 2606.00741 (2026-05-30)
Authors: Uiwon Hwang, Jaeho Hwang
Categories: cs.LG, cs.AI, stat.ML

Problem Statement

As transistors scale to quantum-mechanical limits, thin gate oxides cause electron leakage through quantum tunneling. Unlike conventional digital systems that require perfect correctness, AI inference can tolerate structured errors if properly modeled.

Key Insight

Generic Gaussian noise models miss the structure of quantum tunneling errors. WKB-derived models capture three critical properties:

Exact affine mean drift — systematic bias from tunneling asymmetry
Per-bit variance hierarchy — MSB-dominated error distribution
Per-layer dependence — errors scale with ||W_ℓ||_∞ and network Jacobian

TAC Algorithm: Tunneling-Aware Compensation

Three Structural Properties from WKB Physics

Property	Description	Impact
Affine mean drift	Systematic weight shift from tunneling asymmetry	Closed-form mean correction removes 60%+ of error
Per-bit variance hierarchy	MSB errors dominate over LSB	Optimal bit-budget allocation targets MSB protection
Per-layer dependence	Error scales with weight magnitude & Jacobian	Layer-adaptive protection, not uniform

Algorithm Steps

Mean Correction: Apply closed-form affine correction to each weight matrix
```
W_corrected = W - μ_tunneling(W, oxide_thickness)
```
Bit-Budget Allocation: Distribute ECC budget proportional to WKB-derived variance
```
budget_ℓ ∝ ||W_ℓ||_∞ × Jacobian_sensitivity_ℓ
```
No Retraining Required: Compensation is purely deployment-time

Performance Results

CNNs at p_flip=0.10: 95% of clean accuracy with 3.4×–33.6× less ECC overhead
Transformer at p_flip=0.05: 95% of clean accuracy, massive ECC reduction
Heterogeneous architectures: WKB scoring outperforms magnitude-based by 24pp at small budgets
Zero inference-time overhead: All computation is pre-deployment

Reusable Patterns

Pattern 1: Physics-First Noise Modeling

Instead of assuming Gaussian noise, derive the error distribution from first principles:

Identify the physical mechanism (quantum tunneling in this case)
Apply the appropriate physical approximation (WKB for tunneling)
Extract structural properties from the derived distribution
Design compensation algorithms that exploit the structure

Pattern 2: Saturation Ratio ρ*

The closed-form saturation ratio predicts compensation effectiveness in advance:

ρ* = f(tunneling_rate, weight_magnitude, bit_budget)

Use ρ* to decide whether TAC is worth deploying before running experiments.

Pattern 3: Layer-Adaptive vs. Uniform Protection

Uniform protection (baseline): equal ECC budget across all layers
Layer-adaptive (TAC): budget proportional to WKB-derived sensitivity
At small budgets, layer-adaptive wins by 24 percentage points

Pattern 4: Hardware-Software Co-Design Beyond Scaling

When hardware scaling hits physical limits:

Model the physical error mechanism precisely
Design software compensation that exploits error structure
Trade hardware cost (ECC) for software intelligence (compensation)

Implementation Guidelines

WKB Approximation Parameters:
- Oxide thickness determines tunneling probability
- Work function difference determines asymmetry
- Temperature affects tunneling rate
Jacobian Sensitivity Computation:
- Compute layer-wise Jacobian ∂output/∂W_ℓ
- Sensitivity = ||Jacobian_ℓ||F × ||W_ℓ||∞
- Allocate ECC budget proportional to sensitivity
ECC Budget Optimization:
- Small budgets: prioritize MSB protection on high-sensitivity layers
- Large budgets: extend to LSB protection
- Use ρ* to predict saturation point

Activation Keywords

quantum tunneling, WKB approximation, noise modeling, deployment robustness, hardware-aware ML, error correction, transistor scaling, quantum-mechanical limits, tunneling-aware compensation, TAC, QTAML