PRINT-SAFE: PRINTed ultra-low-cost electronic X-Design with Scalable Adaptive Fault Endurance

This repository contains the code for "PRINT-SAFE: PRINTed ultra-low-cost electronic X-Design with Scalable Adaptive Fault Endurance," published in ESWEEK 2025.

---

Overview

PRINT-SAFE addresses reliability challenges in additive printed electronics (PE) by introducing a novel co-design of training algorithms and hardware for fault-tolerant printed analog neuromorphic circuits (pNCs). Our Fault-Aware Training (FAT) method dynamically selects printed activation functions (AFs) to optimize fault endurance and reduce hardware costs.

---

Key Contributions

• Co-design of Fault-Tolerant pNCs: A pioneering approach integrating circuit and algorithmic design for robust printed electronics.

• Gradient-Based Fault-Aware Training: Utilizes Gumbel-Softmax for differentiable architecture search, adapting to printing defects.

• Improved Fault Tolerance: Achieves significant accuracy improvement (62.1% to 79.4% under 10% fault rate).

• Resource Optimization: Reduces power by 54.5%, area by 6.54%, and training time by 56.2% by combining normal and fault-tolerant AFs.

---

Repository Structure

├── dataset/                # Datasets
├── src/                    # Source code
├── experiment.py           # Main experiment script
├── README.md               # This 

---

SPICE Simulation Data

The Faulty Behavior Dataset (FBDataset) is derived from SPICE simulations of non-linear AFs. Ensure this dataset is available for accurate fault modeling.

---

Running Experiments

Run experiments using experiment.py with command-line arguments:

python3 experiment.py --DATASET 0 --SEED 00 --e_train 0.0 --dropout 0.0 --projectname none_0_FaultAnalysisMixed --type_nonlinear mix --fault_ratio 0.0 --act none

Key Arguments:

• --DATASET: Dataset index.
• --SEED: Random seed.
• --e_train: Training error rate.
• --projectname: Experiment name.
• --type_nonlinear: Type of nonlinearity. Can be normal, robust, or mix.
• --fault_ratio: Fault injection ratio.

Mix Match Implementation

The mix match strategy is primarily implemented in src/pNN_FA_MIX_MATCH.

Citation

Please cite our work if you use this code or concepts:

@article{2025PRINTSAFE,
  author    = {Priyanjana Pal and Tara Gheshlaghi and Haibin Zhao and Michael Hefenbrock and Michael Beigl and Mehdi B. Tahoori},
  title     = {{PRINT-SAFE: PRINTed ultra-low-cost electronic X-Design with Scalable Adaptive Fault Endurance}},
  journal   = {ESWEEK 2025},
}