Research | yaser

Summary of NICS's current research efforts:

1. Investigate New Device Functionalities for Low-Power Neuromorphic Computing Systems based on Emerging Materials, Devices, and Sensors.

Design memristors, transistors, and sensors using emerging materials to construct low-power, ultra-dense, and flexible Neuromorphic ICs.

2. Machine Learning for Solving Domain-specific Real-world Problems and Knowledge Discovery in Science and Engineering

Machine learning techniques are used to create solutions for real-world problems producing automation, classifications, analysis, and predictions.

3. Develop a Device-end to the Algorithm-end Simulator to Design High-performance Computing-in-memory Chips for Neuro-inspired Computing

Develop the device models with hardware non-ideal effects to estimate the performance of the initial neural network architecture for the simulator to change the neural network architecture and parameters accordingly in order to minimize the impact of the non-ideal effects on the AI model accuracy.

4. Develop Brain-inspired Artificial Intelligence Chips for Real-time and Energy-efficient Edge-device Computing

Artificial general intelligence at the edge based on multi-hierarchical efforts

5. Brain-inspired Energy-Efficient and Real-time Sense-and-Avoid Control via Cloud-Edge Framework: Online Supervised Learning with Spiking Neural Networks

Multi-UAV system in the environment with static and dynamic obstacles.

Spiking patterns for each UAV

Inputs provided by the cloud by the edge SNNs implemented on the UAVs for UAV6

Energy consumption for the UAVs in the flock during the mission performance

Accomplished Projects:

Project 1.1. Computational Model of Ta2O5/TaOx Memristors: Predicting Resistive Switching Behavior and Filament Growth Dynamics for Enhanced Device Control and Optimization

Research Area 1. New Device Functionalities for Low-Power Neuromorphic Computing Circuits and Systems based on Emerging Materials, Devices, and Sensors.

Introduce a Ta2O5/TaOx device model that accurately predicts crucial resistive switching properties in memristors, from formation to set/reset cycles, aligning closely with empirical data to enhance device control and optimization.

Introduce a programmable logic gate using circuit implementation of artificial neural network and memristors, exotic electronic devices, only confirmed to exist in 2008.

Project 1.2. Neuromorphic Design: Modeling Memristor as Artificial Synapsis in Neural Network Circuits

Figure: (a) The proposed general SPICE model of hp's memristor in a crossbar ultra dense structure. (b) A two-layer feed forward neural network that has been used as a programmable two input logic gate in this work. (c) Circuit implementation of the neural network-based programmable logic gate implemented by memristors as synapses. (d) Simulation result of the programmable logic gate that shows the training and operation phases for XOR and XNOR logic.

Project 1.3. Performance Evaluation of Graphene Transistors for Low-power Integrated Circuit Design

Study static metrics and switching attributes of graphene nanoribbon field-effect transistor (GNRFETs) using a quantum transport model that self-consistently solves Poisson and Schrödinger equations within the non-equilibrium Green’s function (NEGF) formalism and fully treats on-parabolic band structure, short channel-length electrostatic effects, and the quantum tunneling effects in FET structures.

Figure: (a) Self-consistent calculation between electrostatic (Poisson equation) and transport (NEGF formalism) solutions. (b) Conceptual sketch of the armchair edge GNR channel including the quantities used in the NEGF formalism. (c) Energy-position-resolved local density of states of a typical GNRFET simulated with NEGF formalism, showing thermionic emission of carriers over the channel potential barrier, and the band-to-band tunneling of electrons from valence band in the channel region to the empty states in the drain side. (d) Bandgap energy of three GNR families of armchair GNR versus GNR index and energy dispersion relation of GNR(12,0), GNR(13,0) and GNR(14,0).

Project 1.4. Low-power Tunneling Transistors for Edge Devices and AI Architecture

Design low-power tunnel FETs as the building block of low-power circuits to drive neural networks in neuromorphic architecture.

Figure: 2D materials-based heterostructure tunnel FET (HT-FET) structures that can operate at ultra-low power and ultra-high-speed/frequency. The device functions based on quantum mechanical band-to-band tunneling (BTBT) through a potential barrier. Ultrathin 2D material-based HT-FET decreases tunneling leakage sensing current, increase the sensing current, and reduces power consumption as a result of steep subthreshold swing.

Project 1.5. A physic-based circuit model of graphene nanoribbon FET with edge roughness

Develop a physics-based SPICE-compatible model for GNRFETs and the line-edge roughness of graphene that significantly degrades its performance such as an enhanced leakage current, reduced drive current, and increased threshold voltage.

Figure: (a) 3D view of an all-graphene circuit for an example of inverter chains together with its circuit implementation. The contact resistances via connections are also shown corresponding to the layout design. (b) In the all-graphene circuits, narrow armchair-edge GNRs and wide zigzag-edge GNRs are used as channel material and local interconnects, respectively. (c) 3D view of a GNRFET with one ribbon of armchair GNR(N,0) as channel material. (d) Energy band diagram and the corresponding components in the equivalent circuit model of a GNRFET with one GNR channel. (e) Circuit implementation of a GNR FET with multi-GNR channels consists of five nanoribbons connected to a wide zigzag GNR. (f) IDS–VGS characteristic of GNR FETs with three different GNR indices of N = 6, 12, and 18 for VDS = 0.1V, IDS–VDS characteristics of GNR FET at different gate voltages, and ON- and OFF- currents versus edge roughness amplitude ∆W/W for LCH = 16 nm.

Project 1.6. Power-efficient circuits design based on the clocked adiabatic architecture in graphene FETs

Introduce new energy-efficient gates based on clocked adiabatic logic (CAL) and graphene transistors.

Figure: Circuit schematics of XNOR/XOR and the associated power density versus frequency for (a) 2N-2N2P design, (b) basic CAL design (c) PL-GL CAL design, and (d) Low-Power CAL design.

Project 1.7. Carbon Nanotube Interconnects for Ionized Radiation detection

Explore the feasibility of a carbon nanotube (CNT) ring oscillator (RO) for detecting ionizing radiation using the displacement damage dose (DDD) methodology.

Figure: (a) Schematic of metallic CNT exposed to ionization radiation. (b) Transmission line model of a single-walled CNT interconnect. (c) Change in the ratio of the irradiated resistance, R(DDD) to the pre-irradiation resistance, R(0) as a function of increasing the displacement damage dose. (d) Total resistance versus voltage bias across 10µm length SWCNT (10,10) as a function of increasing DDD from 1012 MeV/g to 1017 MeV/g. The ratio of the irradiated resistance, R(DDD) to the pre-irradiation resistance, R(0) as a function of increasing the displacement damage dose for different (e) lengths and (f) diameter of SWCNT. (g) CNT-based ring oscillator consisting of six CNTFET and three CNT interconnects. Its output waveforms for exposure to ion radiation with three displacement density doses of 1015 MeV/g, 1016 MeV/g, and 1017 MeV/g. (h) Propagation delay of CNT inverter pair versus displacement damage dose for different CNT lengths. (i) Oscillation frequency of CNT-based RO versus DDD for metallic SWCNT (10,10) interconnect with three different lengths of 1 µm, 10 µm, and 100 µm.

Project 1.8. Graphene FET as IoT Temperature Sensors

Introduce and evaluate graphene as a promising material for sensing and monitoring die temperature at scaled bias voltages of emerging technology using a semi-classical transport model incorporating different scattering mechanisms.

Figure: (a) Graphene field-effect transistor (G-FET) as a temperature sensor. (b) Resistance of G-FET versus temperature and gate bias at drain-source of 1V. (c) Temperature coefficient of resistivity (TCR) of the G-FET temperature sensor versus gate bias as a function of drain voltage of different drain-source voltage per Celsius degree variation. (d) Schematic diagram of the back-gate multi-channel graphene nanoribbon (GNR) FET temperature sensor device. (e) Temperature coefficient of resistance (TCR) measurements, illustrate the dependence on GNR width and gate bias.

Project 1.9. Graphene Nanostructure for Controlling Electromagnetic Fields

Introduce new graphene-based aperiodic multilayer structures as selective, tunable, and switchable thermal emitters at infrared frequencies.

Figure: (a) Structure of the proposed thermal emitter composed of alternating layers of graphene and hBN insulator, which are sandwiched between two thick silicon carbide (SiC) layers. (b) Normalized thermal power emitted μ(λ) per unit area and unit wavelength in the normal direction from bulk tungsten versus wavelength and chemical potential at T = 873 K for the five optimized structures with 8, 13,23, 28, and 32 layers of graphene. The optimized graphene-based aperiodic multilayer structures enable narrowband thermal emission. The increase in the chemical potential results in a spectral shift toward shorter wavelength, enabling the electrically tunable thermal emitter, in which the range of tunability increases by increasing the number of graphene layers. The thermal emittance from the optimized structure with a larger number of graphene layers can be almost entirely eliminated by increasing the chemical potential so that perfect emittance of unity for the structure with 23 graphene layers can be switched to zero by setting μc equal to 1.0 eV.

Project 1.10. A Cost-Efficient Surface Enhanced Raman Spectroscopy (SERS) Molecular Detection Technique for Clinical Applications

Develop new techniques for Surface Enhanced Raman Scattering (SERS) without costly and sophisticated fabrication steps.

Figure: Test cells were filled with Rhodamine 6G. Laser light was focused on Au substrate in two cases by a 50 X microscope objective. (a) Au on etched aluminum foil (conventional view), (b) Au on UV epoxy (transmission view). The right figure shows Raman signals for gold on aluminum foil as a function of the concentration of Rhodamine 6G in DI water.

Project 1.11. Resonant tunneling diode for radio-frequency circuits design

Develop a new large-signal SPICE model for RTD and evaluate its applications for a large signal oscillator, an RF amplifier, and a frequency multiplier.

Figure: (a) The SPICE large-signal AC model of resonant tunneling diode (RTD). (b) charge and capacitance versus the applied voltage across the diode. (c) Circuit schematic and the output signal of RTD-based oscillator in time and frequency domains. (d) The input and output voltages of RTD-based amplifier at an optimum bias condition.

Project 1.12. Design of Microresonators/Detectors to Minimize Thermal Noise Below Standard Quantum Limit

Develop a new microresonator detector and optimize its geometery to minimize its thermal noise below standard quantum limit and to evaluate its applications for a LIGO gravitational wave detection.

Figure: (a) Experimental arrangement of our optomechanical system and the optomechanical cavity consists of a macroscopic end mirror and a microresonator. (b) Thermal noise and standard quantum limit as a function of frequency for the microresonators, and the SQL/TN ratio for the microresonators, as well as the definition of the parameters. (c) RMAX and (d) BWE as a function of the lengths and widths of the optimized microresonators.

Research Area 2. Machine Learning Application for Solving Domain-specific Real-world Problems and Knowledge Discovery in Sciences and Engineering

Project 2.1. Harnessing Smart Additive Manufacturing using Machine Learning Techniques

Develop a closed-loop machine learning algorithm and improve the underlying failure phenomena in laser-based 3D metal printing.

Figure: (a) Two-step strategy to establish the closed-loop machine learning algorithm for additive manufacturing process control; (b) Schematics of the objects in design software, corresponding printed parts with substantial defects on their surfaces that are detected and labeled by the CNN model as “defected” or “Not defected"; (c) Convolutional neural network procedure that extracts spatial features to pass to aggregation layers (averaging, pooling, etc.), followed by higher-order features extraction that is combined at the top layer for fault detection and classification.

Project 2.2. Identifying Laguerre-Gaussian Modes using Convolutional Neural Network

Develop an automated determination of Laguerre-Gaussian (LG) modes to improve cavity tuning and optical communication.

Figure: Training dataset of first 16 LG modes (l and p vary from 0 to 3) are passed to the CNN model that extracts spatial features using aggregation layers (averaging, pooling, etc.) and additional filter layers that are eventually combined at the top layer to predict and classify the new image of LG mode. Prediction confidence on two experimental setup of LG modes are also shown. The dark red and dark blue indicate the results of the CNN models trained using the simulated images while the light red and light blue indicate the results of the CNN models trained using both the simulated and experiment images.

Project 2.3. Enabling Fully Automated Analysis of Microscopy Images using CNN

Develop an accurate, fast and fully automated cell segmentation, classification, and counting techniques for large-scale high-throughput microscopy images.

Figure: Preparation of samples and data sets of microscopy images, annotations by manually drawing bounding rectangles around cells associated with macrophages (identify polarized macrophages M1 and M2 phenotypes and distinguish them from unpolarized macrophages and monocytes), training CNN, and finally evaluation of the accuracy and performance of the framework by calculating precision, recall and F1 score of the cell counts and classifications to compare with the results of human experts and two other open-source software tools.

Project 2.4. Data-driven Analysis of Gamma-Ray Flashes Produced by Lightning Observed at Ground Level

Develop a data collection and interpretation algorithm to record bursts of gamma-rays of millisecond-scale duration associated with lightning strikes, analyze the pattern of events, and predict the thunderstorm characteristics.

Figure: Radio/optical systems for observing terrestrial lightning flashes

Project 2.5. Text-Mining and Image Processing Techniques for Generating Big Geological and Hydrological Data Sets

Data preprocessing of bulky handwritten geotechnical data, gathered over the decades in Southeastern US using OCR and image processing technology to prepare structured big geological and hydrological data sets.

Figure: The procedure to convert the handwritten soil/sand data from well logs into structural machine-readable format

Project 2.6. Detecting Malicious DNS over HTTPS Traffic in Domain Name System using Machine Learning Classifiers

Develop a systematic two-layer approach for detecting DNS over HTTPS (DoH) traffic and distinguishing Benign-DoH traffic from Malicious-DoH traffic using six machine learning algorithms.

Figure: (a) Training procedure of intrusion detection including data preprocessing, training, and optimizing the training algorithms, deployment of ML-based classifiers, and testing of the model to extract the classification performance metrics. (b) The network topology used to capture the traffic datasets, including benign and malicious DoH traffic along with non-DoH traffic.

Project 2.7. Design of Network Intrusion Detection Systems for the Internet of Things Infrastructure using Machine Learning Algorithms

Evaluating the performance of machine learning algorithms for network intrusion detection systems on the internet of things infrastructure.

Figure: (a) Training progress pipelines for four ML models: XGBoost classifier, random forest classifier, decision tree classifier, and gradient boosting classifier; (b) Classification measures of XGB classifier for four different sizes of attack datasets.