Search

Reliable uncertainty estimation plays a crucial role in various safety-critical applications such as medical diagnosis and autonomous driving. In recent years, Bayesian neural networks (BayesNNs) have gained substantial research and industrial interests due to their capability to make accurate predictions with reliable uncertainty estimation. However, the algorithmic complexity and the resulting hardware performance of BayesNNs hinder their adoption in real-life applications. To bridge this gap, this paper proposes an algorithm and hardware co-design framework that can generate field-programmable gate array (FPGA)-based accelerators for efficient BayesNNs. At the algorithm level, we propose novel multi-exit dropout-based BayesNNs with reduced computational and memory overheads while achieving high accuracy and quality of uncertainty estimation. At the hardware level, this paper introduces a transformation framework that can generate FPGA-based accelerators for the proposed efficient multi-exit BayesNNs. Several optimization techniques such as the mix of spatial and temporal mappings are introduced to reduce resource consumption and improve the overall hardware performance. Comprehensive experiments demonstrate that our approach can achieve higher energy efficiency compared to CPU, GPU, and other state-of-the-art hardware implementations. To support the future development of this research, we have open-sourced our code at: https://github.com/os-hxfan/MCME_FPGA_Acc.git, Comment: arXiv admin note: text overlap with arXiv:2308.06849

We characterize molecular magnet heterojunctions in which sublimated CoPc films as thin as 5 nm are sandwiched between transparent conducting bottom-layer indium tin oxide and top-layer soft-landing eutectic GaIn (EGaIn) electrodes. The roughness of the cobalt phthalocyanine (CoPc) films was determined by atomic force microscopy to be on the order of several nanometers, and crystalline ordering of lying-down planar molecules was confirmed by X-ray diffraction. The current-voltage (I-V) characteristics reveal the onset of a superconducting gap at Tc = 6 K, which together with higher temperature fits to a modified Simmons' model, provide incontrovertible evidence for direct quantum mechanical tunneling processes through the magnetic molecules in our heterojunctions. The voltage dependent features in the differential conductance measurements relate to spin states of single molecules or aggregates of molecules and should prove to be important for quantum information device development., Comment: 14 pages, 5 figures

In certain configurations, the aromatic properties of benzene ring structured molecules allow for unpaired, reactive valence electrons (known as radicals). Clar's goblets are such molecules. With an even number of unpaired radicals, these nanographenes are topologically frustrated hydrocarbons in which pi-bonding network and topology of edges give rise to the magnetism. Clar's goblets are therefore valued as prospective qubits provided they can be modulated between magnetic states. Using first principles DFT, we demonstrate the effects of adsorption on both molecule and substrate in a graphene-Clar's goblet heterostructure. We look at the energy difference bewteen FM and AFM states of the system and discuss underlying physical and chemical mechanisms in reference to the highest occupied molecular orbital (HOMO) and second HOMO (HOMO-1). We find that the HOMO of the molecule in the FM state is right at the Fermi surface, which leads to the hybridization between molecular state and the graphene state near the Dirac point. Furthermore, we investigate qualitative changes in charge realignment and magnetic state under variable electric field. Transitions from FM to AFM and back to FM states are observed., Comment: 10 pages, 8 figures

\ce{MoS_2} is a two dimensional material with a band gap depending on the number of layers and tunable by an external electric field. The experimentally observed intralayer band-to-band tunneling and interlayer band-to-band tunneling in this material present an opportunity for new electronic applications in tunnel field effect transistors. However, such a widely accepted concept has never been supported up by theoretical investigations based on first principles. In this work, using density functional theory, in conjunction with non-equilibrilibrium Green's function techniques and our electric field gating method, enabled by a large-scale computational approach, we study the relation between band alignment and transmission in planar and side-stack \ce{MoS_2} $p$-$i$-$n$ junction configurations. We demonstrate the presence of negative differential resistance for both in-plane and interlayer current, a staple characteristic of tunnel diode junctions, and analyze the physical origin of such an effect. Electrostatic potentials, the van der Waals barrier, and complex band analysis are also examined for a thorough understanding of Esaki Diodes., Comment: 8 pages, 8 figures

Recent advances in algorithm-hardware co-design for deep neural networks (DNNs) have demonstrated their potential in automatically designing neural architectures and hardware designs. Nevertheless, it is still a challenging optimization problem due to the expensive training cost and the time-consuming hardware implementation, which makes the exploration on the vast design space of neural architecture and hardware design intractable. In this paper, we demonstrate that our proposed approach is capable of locating designs on the Pareto frontier. This capability is enabled by a novel three-phase co-design framework, with the following new features: (a) decoupling DNN training from the design space exploration of hardware architecture and neural architecture, (b) providing a hardware-friendly neural architecture space by considering hardware characteristics in constructing the search cells, (c) adopting Gaussian process to predict accuracy, latency and power consumption to avoid time-consuming synthesis and place-and-route processes. In comparison with the manually-designed ResNet101, InceptionV2 and MobileNetV2, we can achieve up to 5% higher accuracy with up to 3x speed up on the ImageNet dataset. Compared with other state-of-the-art co-design frameworks, our found network and hardware configuration can achieve 2% ~ 6% higher accuracy, 2x ~ 26x smaller latency and 8.5x higher energy efficiency.

In this computational work based on density functional theory we study the electronic and electron transport properties of asymmetric multi-layer MoSSe junctions, known as Janus junctions. Focusing on 4-layer systems, we investigate the influence of electric field, electrostatic doping, strain, and interlayer stacking on the electronic structure. We discover that a metal to semiconductor transition can be induced by an out-of-plane electric field. The critical electric field for such a transition can be reduced by in-plane biaxial compressive strain. Due to an intrinsic electric field, a 4-layer MoSSe can rectify out-of-plane electric current. The rectifying ratio reaches 34.1 in a model junction Zr/4-layer MoSSe/Zr. This ratio can be further enhanced by increasing the number of MoSSe layers. In addition, we show a drastic sudden vertical compression of 4-layer MoSSe due to in-plane biaxial tensile strain, indicating a second phase transition. Furthermore, an odd-even effect on electron transmission at the Fermi energy for Zr/$n$-layer MoSSe/Zr junctions with $n=1, \, 2,\, 3, \,\dots,\, 10$ is observed. These findings reveal the richness of physics in this asymmetric system and strongly suggest that the properties of 4-layer MoSSe are highly tunable, thus providing a guide to future experiments relating materials research and nanoelectronics.

We study gate field effects on the Mn$_{12}$O$_{12}$(COOH)$_{16}$(H$_2$O)$_4$ | graphene | GaAs heterostructure via first-principles calculations. We find that under moderate doping levels electrons can be added to but not taken from the single-molecule magnet Mn$_{12}$O$_{12}$(COOH)$_{16}$(H$_2$O)$_4$ (Mn$_{12}$). The magnetic anisotropy energy (MAE) of Mn$_{12}$ decreases as the electron doping level increases, due to electron transfer from graphene to Mn$_{12}$ and change in the band alignment between Mn$_{12}$ and graphene. At an electron doping level of $-5.00 \times 10^{13}\, \textrm{cm}^{-2}$, the MAE decreases by about 18% compared with zero doping. The band alignment between graphene and GaAs is more sensitive to electron doping than to hole doping since the valence band of GaAs is close to the Fermi level. The GaAs substrate induces a small bandgap in the supported graphene under the zero gate field and a nearly strain-free configuration. Finally, we propose a vertical tunnel junction for probing the gate dependence of MAE via electron transport measurements.

This review article provides a bird's-eye view of what first-principles based methods can contribute to next-generation device design and simulation. After a brief overview of methods and capabilities in the area, we focus on published work by our group since 2015 and current work on $\textrm{CrI}_3$. We introduce both single- and dual-gate models in the framework of density functional theory and the constrained random phase approximation in estimating the Hubbard $U$ for 2D systems vs. their 3D counterparts. A wide range of systems, including graphene-based heterogeneous systems, transition metal dichalcogenides, and topological insulators, and a rich array of physical phenomena, including the macroscopic origin of polarization, field effects on magnetic order, interface state resonance induced peak in transmission coefficients, spin filtration, etc., are covered. For $\textrm{CrI}_3$ we present our new results on bilayer systems such as the interplay between stacking and magnetic order, pressure dependence, and electric field induced magnetic phase transitions. We find that a bare bilayer $\textrm{CrI}_3$, graphene$\,|\,$bilayer $\textrm{CrI}_3\,|\,$graphene, $h$-BN$\,|\,$bilayer $\textrm{CrI}_3\,|\,h$-BN, and $h$-BN$\,|\,$bilayer $\textrm{CrI}_3\,|\,$graphene all have a different response at high field, while small field the difference is small except for graphene$\,|\,$bilayer $\textrm{CrI}_3\,|\,$graphene. We conclude with discussion of some ongoing work and work planned in the near future, with the inclusion of further method development and applications.

Magic-angle twisted bilayer graphene (MATBG) is notable as a highly tunable platform for investigating strongly correlated phenomena such as high-$T_c$ superconductivity and quantum spin liquids, due to easy control of doping level through gating and sensitive dependence of the magic angle on hydrostatic pressure. Experimental observations of correlated insulating states, unconventional superconductivity and ferromagnetism in MATBG indicate that this system exhibits rich exotic phases. In this work, using density functional theory calculations in conjunction with the effective screening medium method, we find the MATBG under pressure at a twisting angle of $2.88\unicode{xb0}$ and simulate how its electronic states evolve when doping level and out-of-plane electric field are gate-tuned. Our calculations show that, at doping levels between two electrons and four holes per moir\'{e} unit cell, a ferromagnetic solution with spin density localized at AA stacking sites is lower in energy than the nonmagnetic solution. The magnetic moment of this ferromagnetic state decreases with both electron and hole doping and vanishes at four electrons/holes doped per moir\'{e} unit cell. Hybridization between the flat bands at the Fermi level and the surrounding dispersive bands can take place at finite doping. Moreover, upon increasing the out-of-plane electric field at zero doping, a transition from the ferromagnetic state to the nonmagnetic one is seen. We also analyze the interlayer bonding character due to the flat bands via Wannier functions. Finally, we report trivial band topology of the flat bands in the ferromagnetic state at a certain doping level.

We use density functional theory to study the structural, magnetic and electronic structure of the organo-metallic quantum magnet $\mathrm{NiCl_2-4SC(NH_2)_2}$ (DTN). Recent work has demonstrated the quasi-1D nature of the molecular crystal and its quantum phase transitions at low temperatures. This includes a magneto-electric coupling and, when doped with Br, the presence of an exotic Bose-glass state. We systematically show that, by using the generalized gradient approximation (GGA) with inclusion of a van der Waals term to account for weak inter-molecular forces and by introducing a Hubbard $U$ term to the total energy, our calculations reproduce the magnetic anisotropy, the inter-molecular exchange coupling strength and the magneto-electric effect in DTN, which were observed in previous experiments. Further analysis into the electronic structure gives insight into the underlying magnetic interactions, including what mechanisms may be causing the ME effect. Using this computationally efficient model, we predict what effect applying an electric field might have on the magnetic properties of this quantum magnet.

Interfaces between two topological insulators are of fundamental interest in condensed matter physics. Inspired by experimental efforts, we study interfacial processes between two slabs of BiSbTeSe2 (BSTS) via first principles calculations. Topological surface states are absent for the BSTS interface at its equilibrium separation, but our calculations show that they appear if the inter-slab distance is greater than 6 Ang. More importantly, we find that topological interface states can be preserved by inserting two or more layers of hexagonal boron nitride between the two BSTS slabs. In experiments, the electric current tunneling through the interface is insensitive to back gate voltage when the bias voltage is small. Using a first-principles based method that allows us to simulate gate field, we show that at low bias the extra charge induced by a gate voltage resides on the surface that is closest to the gate electrode, leaving the interface almost undoped. This provides clues to understand the origin of the observed insensitivity of transport properties to back voltage at low bias. Our study resolves a few questions raised in experiment, which does not yet offer a clear correlation between microscopic physics and transport data. We provide a road map for the design of vertical tunneling junctions involving the interface between two topological insulators.

Hyperspectral image (HSI) classification has been widely adopted in applications involving remote sensing imagery analysis which require high classification accuracy and real-time processing speed. Methods based on Convolutional neural networks (CNNs) have been proven to achieve state-of-the-art accuracy in classifying HSIs. However, CNN models are often too computationally intensive to achieve real-time response due to the high dimensional nature of HSI, compared to traditional methods such as Support Vector Machines (SVMs). Besides, previous CNN models used in HSI are not specially designed for efficient implementation on embedded devices such as FPGAs. This paper proposes a novel CNN-based algorithm for HSI classification which takes into account hardware efficiency. A customized architecture which enables the proposed algorithm to be mapped effectively onto FPGA resources is then proposed to support real-time on-board classification with low power consumption. Implementation results show that our proposed accelerator on a Xilinx Zynq 706 FPGA board achieves more than 70x faster than an Intel 8-core Xeon CPU and 3x faster than an NVIDIA GeForce 1080 GPU. Compared to previous SVM-based FPGA accelerators, we achieve comparable processing speed but provide a much higher classification accuracy., Comment: This article is accepted for publication at ARC'2019

By solving two-component spinor equation for massless Dirac Fermions, we show that graphene under a periodic external magnetic field exhibits a unique energy spectrum: At low energies, Dirac Fermions are localized inside the magnetic region with discrete Landau energy levels, while at higher energies, Dirac Fermions are mainly found in non-magnetic regions with continuous energy bands originating from wavefunctions analogous to particle-in-box states of electrons. These findings offer a new methodology for the control and tuning of massless Dirac Fermions in graphene., Comment: 9 pages, 4 figures

We show that when a molecular junction is under an external bias, its properties can not be uniquely determined by the total electron density in the same manner as the density functional theory (DFT) for ground state (GS) properties. In order to correctly incorporate bias-induced nonequilibrium effects, we present a dual mean field (DMF) approach. The key idea is that the total electron density together with the density of current-carrying electrons are sufficient to determine the properties of the system. Two mean fields, one for current-carrying electrons and the other one for equilibrium electrons can then be derived.By generalizing the Thomas-Fermi-Dirac (TFD) model to non-equilibrium cases, we analytically derived the DMF exchange energy density functional. We implemented the DMF approach into the computational package SIESTA to study non-equilibrium electron transport through molecular junctions. Calculations for a graphene nanoribbon (GNR) junction show that compared with the commonly used \textit{ab initio} transport theory, the DMF approach could significantly reduce the electric current at low biases due to the non-equilibrium corrections to the mean field potential in the scattering region., Comment: 5 pages, 3 figures