Machine Learning-based model for Single Event Upset Current Prediction in 14nm FinFETs

This work presents a machine learning regression-based surrogate model of Single Event Upset (SEU) transient current for circuit-level simulation. The phenomenal success of FinFET technology in terms of integration and performance over planar MOSFETs has paved the way for their usage in aerospace-integrated circuits and defense applications. However, their sensitivity to radiation hazards in such applications remains the primary concern. With the recent technological advancement, the semiconductor industry has shifted its focus to device analysis before fabrication so that the circuit designers may mitigate radiation effects before actual fabrication. The Technology Computer-Aided Design (TCAD) tools are being used to design the structure and analyze the device parameters. However, these tools are computationally intensive and time-consuming. This work explores the feasibility of using machine learning for predicting device parameters and Single Event Transient (SET) current using an unsupervised learning technique. A 14nm 3D FinFET device is designed using the TCAD tool, and a dataset with various parameters is generated. This dataset is used to train (1) a Random Forest Regressor model and (2) A feedforward neural network for predicting SET pulse current. The 10% dataset was randomly chosen as a subset to test this algorithm and predict SET current. The comparison between actual and predicted data shows high accuracy. For example, the random forest algorithm achieves a mean square error of 1.49e-3 for the test dataset. This shows that machine learning models can replace TCAD for accelerating device performance analysis for large-scale circuits. The conventional TCAD simulation takes 4 hours per simulation on a Xeon W1350P processor and 32 GB RAM hardware. By contrast, our proposed model takes only 8–10 seconds to predict the SET current. This study can help designers mitigate SET effects in the design phase. The source-code of our proposed machine-learning models is available at https://github.com/vihhu53/MLSEUFinfet.