Fully Convolutional Generative Machine Learning Method for Accelerating Non-Equilibrium Green’s Function Simulations

Aug 3, 2026

Upload by Finn

A novel simulation workflow combines convolutional generative machine learning with quantum-mechanical non-equilibrium Green’s function (NEGF) device modeling. The method, ML-NEGF, is implemented in the NESS (nano-electronics simulations software) in-house simulator and targets faster convergence for coupled Poisson–NEGF computations. Training enables the model to learn the underlying physics of nano-sheet transistor behavior while maintaining accuracy. Results show an average convergence acceleration of 60%, substantially reducing computational time compared with standard NEGF.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

Introduction

Device Structure

Simulations Methodology

Fully Convolutional Generative Machine Learning Method for Accelerating Non-Equilibrium Green’s Function Simulations Preslav Aleksandrov Device Modelling Group James Watt School of Engineering University of Glasgow, UK [preslav.aleksandrov@glasgow.ac.uk](preslav.aleksandrov@glasgow.ac.uk) Tapas Dutta Device Modelling Group James Watt School of Engineering University of Glasgow, UK [tapas.dutta@glasgow.ac.uk](tapas.dutta@glasgow.ac.uk) Ali Rezaei Device Modelling Group James Watt School of Engineering University of Glasgow, UK [ali.reazaei@glasgow.ac.uk](ali.reazaei@glasgow.ac.uk) Asen Asenov Device Modelling Group James Watt School of Engineering University of Glasgow, UK[asen.asenov@glasgow.ac.uk](asen.asenov@glasgow.ac.uk) Nikolas Xeni Device Modelling Group James Watt School of Engineering University of Glasgow, UK [nikolas.xeni@glasgow.ac.uk](nikolas.xeni@glasgow.ac.uk) Vihar Georgiev Device Modelling Group James Watt School of Engineering University of Glasgow, UK [vihar.georgiev@glasgow.ac.uk](vihar.georgiev@glasgow.ac.uk) Abstract—This work describes a novel simulation approach that combines machine learning and device modeling simulations. The device simulations are based on the quantum mechanical non-equilibrium Green’s function (NEGF) approach and the machine learning method is an extension to a convolutional generative network. We have named our new simulation approach ML-NEGF and we have implemented it in our in-house simulator called NESS (nano-electronics simulations software). The reported results demonstrate the improved convergence speed of the ML-NEGF method in comparison to the ‘standard’ NEGF approach. The trained ML model effectively learns the underlying physics of nano-sheet transistor behaviour, resulting in faster convergence of the coupled Poisson-NEGF simulations. Quantitatively, our MLNEGF approach achieves an average convergence acceleration of 60%, substantially reducing the computational time while maintaining the same accuracy. Keywords— machine learning, neural network, autoencoder, device simulations and modeling, non-equilibrium Green’s function (NEGF), TCAD device modeling, nanowires. I. INTRODUCTION The silicon nanowire and nanosheet transistors have a wide spectrum of promising applications [1], such as current field-effect transistors [2] and photovoltaics [3] . Moreover, the state-of-the-art CMOS technologies are based on single or stacked configurations of nanosheet or nanowire architectures [4] . However, despite the recent advances in technology, there is still room to improve the fabrication process and to optimise device performance, for example, by reducing power consumption and reducing device-to-device variability during the fabrication process. From a practical point of view, simulations and modelling transistors are the most time-efficient and cost-effective approach to evaluate the performance and the output characteristics of transistors. The aim is to have a simulation platform that is fast, accurate, and reliable in order to aid the improvement of device design, predict device performance (current-voltage characteristics) and extract important Figures of Merit (FoM) . The main aim of this work is to investigate the possibility of significantly improving or even replacing numerical Fig. 1. Diagram of the n-type silicon (Si) nanosheet transistor with channel cross section of 3 nm x 12 nm (a) and length of full device 22 nm (b). Gate length is 16 nm and the source/drain have length of 3 nm each. The channel doping is 1e16 cm-3 and the contacts (source and drain have) are 1e20 cm-3. The oxide is SiO2 with thickness of 1nm everywhere around the device. Technology Computer-Aided Design (TCAD) device simulations with a convolutional autoencoder (CAE) [5] [6] [7] . To test our idea, we have developed a new simulation approach based on the combination of TCAD and machine learning methods. The current state of the art of TCAD simulations is based on th

FAQs

What problem does ML-NEGF address in NEGF device simulations?

ML-NEGF improves the convergence speed of NEGF-based coupled Poisson–NEGF simulations, reducing overall computational time while keeping accuracy.

What underlying simulation and learning components does the method combine?

The device simulation relies on the quantum-mechanical non-equilibrium Green’s function (NEGF) formalism, while the learning component extends a convolutional generative network (convolutional autoencoder-based approach).

How much convergence acceleration is reported compared with standard NEGF?

The ML-NEGF approach achieves an average convergence acceleration of 60% versus the standard NEGF approach.

Aug 3, 2026

Upload by Finn