TL;DR
This paper introduces a novel machine learning approach to model Single Event Transients (SETs) in integrated circuits. By training a neural network on TCAD simulation data, they achieve accurate SET current models that can be integrated into circuit simulators, offering a faster and more practical alternative to complex physics-based models.
What Problem Does This Paper Address?
As integrated circuits shrink and operating voltages decrease, they become increasingly susceptible to Single Event Transients (SETs) caused by radiation. Accurately simulating these SETs is crucial for designing reliable aerospace and other radiation-hardened electronics. Traditional methods for modeling SETs either rely on complex, computationally expensive physics-based models that are difficult to integrate into circuit-level simulations, or use simplified functional models that struggle to capture the intricate interplay of various physical parameters like Linear Energy Transfer (LET), strike location, and bias voltages. This paper aims to bridge this gap by developing an accurate and computationally efficient SET model suitable for circuit-level simulation.
Methods
The core of this paper is the application of a machine learning regression technique, specifically a multi-layer feedforward neural network, to model SET current pulses. The process involves:
- TCAD Simulation: Generating a comprehensive dataset of SET current pulses using Cogenda VisualTCAD. This simulation captures the device behavior under varying conditions, including different LET values, drain bias voltages, and over time.
- Data Pre-processing: The raw TCAD data is pre-processed. Cubic spline interpolation is used to obtain sampling points, particularly at curve corners, to prevent neural network overfitting. The data is then split into training, validation, and testing sets (70%, 15%, 15% respectively).
- Neural Network Training: A feedforward neural network is trained using the Levenberg-Marquardt algorithm. The network architecture (number of layers, neurons, and transfer functions) is optimized by comparing the Mean Squared Error (MSE) of various configurations. The chosen architecture is an 8x8x1 network.
- Model Implementation: The trained neural network model is implemented as a Verilog-A current source in the Cadence Spectre circuit simulator.
Datasets
The dataset used for training and evaluation was generated through TCAD simulations using Cogenda VisualTCAD. The simulations covered:
- Device: SMIC 130nm bulk CMOS NMOS transistor.
- Parameters varied:
- Linear Energy Transfer (LET): 4 to 100 MeV·cm²/mg.
- Drain Bias Voltage (V<sub>D</sub>): 0 to 1.8V.
- Time: Simulated over the duration of the SET pulse.
- Total Data Points: 418,000 data points were generated after interpolation.
Key Insights and Ideas
- Machine Learning for SET Modeling: The paper demonstrates that machine learning, specifically neural networks, can effectively learn the complex, non-linear behavior of SET currents without explicitly modeling the underlying physics.
- Accuracy vs. Complexity Trade-off: The authors systematically explored different neural network architectures to find an optimal balance between model accuracy and computational complexity, a critical factor for circuit-level simulations.
- Practical Circuit-Level Integration: By implementing the trained model as a Verilog-A current source, the authors show a clear path for integrating this ML-based model into standard circuit simulators, making it readily usable by designers.
- Importance of Pulse Characteristics: The trained model accurately captures key SET pulse characteristics like peak value and width, which are vital for understanding circuit-level transient behavior.
Key Results
- The proposed machine learning regression model achieved a good fit to the SET pulse current data generated by TCAD simulations, particularly for peak value and pulse width.
- The trained SET pulse current model was successfully implemented as a Verilog-A current source in Cadence Spectre.
- A simulation of a two-stage CMOS inverter using the Verilog-A SET current source demonstrated the practicability and reasonableness of the proposed model in predicting circuit-level SET behavior. The simulation showed the output voltage being pulled down from a logic high state due to the ion strike.
Limitations and Future Work
- Limitations: The paper focuses on NMOS transistors and a specific technology node (SMIC 130nm CMOS). The accuracy of the model is dependent on the quality and coverage of the TCAD simulation data. The paper does not explicitly discuss the computational cost of the training process itself, only the inference speed.
- Future Work: The authors suggest that the analysis can be extended to other digital or analog circuits using the proposed SET pulse current model. This implies potential for broader application across different circuit designs and potentially other device types. Further work could also explore more advanced ML architectures or techniques for even higher accuracy or faster training.
Summary updated