Deep learning-enabled prediction of 2D material breakdown

Abstract

Characterizing electrical breakdown limits of materials is a crucial step in device development. However, methods for repeatable measurements are scarce in two-dimensional materials, where breakdown studies have been limited to destructive methods. This restricts our ability to fully account for variability in local electronic properties induced by surface contaminants and the fabrication process. To tackle this, we implement a two-step deep-learning model to predict the breakdown mechanism and breakdown voltage of monolayer MoS2 devices with varying channel lengths and resistances using current measured in the low-voltage regime as inputs. A deep neural network (DNN) first classifies between Joule and avalanche breakdown mechanisms using partial current traces from 0 to 20 V. Following this, a convolutional long short-term memory network (CLSTM) predicts breakdown voltages of these classified devices based on partial current traces. We test our model with electrical measurements collected using feedback-control of the applied voltage to prevent device destruction, and show that the DNN classifier achieves an accuracy of 79% while the CLSTM model has a 12% error when requiring only 80% of the current trace as inputs. Our results indicate that information encoded in the current behavior far from the breakdown point can be used for breakdown predictions, which will enable non-destructive and rapid material characterization for 2D material device development.

A systematic comparison of the structural and dynamic properties of commonly used water models for molecular dynamics simulations

“…Finally, we used machine learning methods to explore the contributions of the parameters (model type, partial charge, Lennard-Jones parameters, bond length and angle) to the surface tension, dielectric constant, and self-diffusion coefficient. Machine learning can be used to predict a wide variety of material, [86][87][88][89] chemical, [90][91][92][93] and biological properties, [94][95][96][97] with several commercial and non-commercial open-source platforms that can be used to develop machine learning algorithms such as Schrödinger, 98 SYBYL, 99 TensorFlow (Google), 100 and BioPPSy. 101…”

Systematic Comparison of the Structural and Dynamic Properties of Commonly Used Water Models for Molecular Dynamics Simulations

“…Finally, we used machine learning methods to explore the contributions of the parameters (model type, partial charge, Lennard-Jones parameters, bond length, and angle) to the surface tension, dielectric constant, and self-diffusion coefficient. Machine learning can be used to predict a wide variety of material, − chemical, − and biological properties, − with several commercial and noncommercial open-source platforms that can be used to develop machine learning algorithms such as Schrödinger, SYBYL, TensorFlow (Google), and BioPPSy …”

Dielectrics for Two-Dimensional Transition-Metal Dichalcogenide Applications

“…Small EOTs lead to devices with efficient gate control and small subthreshold swings, i.e., how quickly a transistor turns on and off, that can approach the room temperature thermionic limit of 60 mV/dec. Resulting high current on/off ratios and low operating voltages increase energy efficiency, while high breakdowns fields contribute toward reliable operation over a typical lifetime of 10 years. ,,, …”