The invention relates to a synaptic component for a neural network. By synaptic component is meant an electronic component having properties of a biological synapse. In the following, the electronic component is therefore also referred to as a synapse. By neural network is meant an artificial neural network.
A brain is a biological neural network. It comprises a plurality of biological neurons, i.e. nerve cells. A biological neuron can transmit information to another biological neuron via biological synapses. This is done by means of an output channel of the biological neuron. Such an output channel is called an axon. The other biological neuron has an input channel. The input channel is called a dendrite. Through the input channel, the other biological neuron can receive the information. The decision of whether a biological neuron triggers an action potential through its axon and thus transmits information depends on the totality of incoming signals. The influence of one biological neuron on another biological neuron om biological synapses can change over time with the activity of the biological synapses. Such influence is called “synaptic weight” of a biological neuron.
The properties and behaviors of a neural network should be similar to the properties and behaviors of a biological neural network. A neural network shall be able to recognize patterns by means of adaptive algorithms. A neural network shall, comparable to the human brain, not have to start from scratch for every task or problem. It should have the possibility to draw on already acquired knowledge and already made experiences.
A neural network with memristors is known from the publication CN 208922326 U. Scaling, stability and blocking capability are problematic for a neural network with memristors.
The publication WO 2019147859 A2 discloses an electronic component having a semiconductor channel. An input electrode is adjacent to one end of the semiconductor channel. An output electrode is adjacent to the other end of the semiconductor channel. Above the semiconductor channel, there is a layer of dielectric material. A circuit of semiconductor materials serving as a synapse with an electrically charged carrier gas is described in the publication EP 0 529 565 B1. An integrated circuit for providing a synapse is known from the publication US 2019164597 A1.
If a neural network is to be realized with conventional CMOS technology, a plurality of transistors is required. Manufacturing effort and energy consumption are then high. A neural network can comprise ferroelectric field effect transistors (FeFET). Traditional ferroelectric materials such as lead zirconate titanate (PZT) are harmful to the environment and incompatible with established CMOS technology.
It is the task of the invention to provide a synapse, i.e., a synaptic component with improved properties.
A synaptic component (synaptic device) having the features of claim 1 serves to solve the task. The dependent claims relate to advantageous embodiments of the invention. The additional claim relates to a method for operating the synaptic component.
Claim 1 relates to a synaptic component for a neural network. The synaptic component comprises a layer consisting of a semiconductor. The layer consisting of a semiconductor is called a semiconducting layer. A source electrode is connected to the semiconducting layer. A drain electrode is connected to the semiconducting layer. The source electrode is spatially separated from the drain electrode. The source electrode and the semiconducting layer form a Schottky diode. The source electrode is separated from a first gate electrode by ferroelectric material. The drain electrode may be separated from a second gate electrode by ferroelectric material. The two gate electrodes are spatially separated from each other.
An electrode is separated from another electrode if there is a distance between the two electrodes. Spatially separated electrodes therefore do not contact each other. There is therefore no electrically conductive connection between the two electrodes. By means of the ferroelectric material, an electrical voltage applied to the first gate electrode can be transferred to the source electrode. The same applies in the case that a second gate electrode is separated from the drain electrode by ferroelectric material. Usually, one gate electrode is located on one side of the ferroelectric material and the other source and/or drain electrode is located on an opposite side of the ferroelectric material. The two electrodes may be directly connected to the ferroelectric material. The invention exploits the memory properties of ferroelectric material to imitate the behavior of a biological synapse.
A Schottky diode may comprise a metal layer applied on a semiconducting layer such as a silicon layer. A metal layer is a layer that consists of metal. A silicon layer is a layer that consists of silicon. Silicon is a semiconductor, i.e. a chemical substance whose electrical conductivity is between that of electrical conductors and that of electrical insulators.
The silicon layer can be, for example, n silicon, i.e. an n-type doped silicon layer. Electrons from the n-conductive silicon layer migrate to the metal layer. Because electrons move more easily from n-conductive silicon to the metal layer than vice versa, a region depleted of electrons is formed in the silicon layer. This region is called a Schottky barrier. Thus, a barrier layer and/or space charge zone is created. In addition, an electric field is created. The electric field counteracts the migration of the electrodes. If the electric field is large enough, electrons no longer migrate from the n-conductive silicon layer to the metal layer. However, the silicon layer can also be p silicon, i.e. a p-type doped silicon layer.
If a positive electrical voltage is applied to n-conductive silicon and a negative electrical voltage is applied to the metal layer, the space charge zone becomes larger. The Schottky diode is then connected in reverse direction.
If an electrically negative voltage is applied to n-conductive silicon and an electrically positive voltage is applied to the metal layer, then the Schottky diode is connected in the forward direction. Electrons flow from n silicon into the metal layer. If the Schottky diode is connected in forward direction, then the space charge zone is cleared.
An electrode is an electrical conductor. An electrode generally consists of metal.
A layer according to the invention may run in a straight line, i.e. along a plane. However, a layer according to the invention can also be round and/or angular.
With the component having the features of claim 1, the behavior of a biological synapse can be imitated. The component can be manufactured with low technical effort. The component can be manufactured from environmentally friendly materials.
The source electrode and the drain electrode may be on one side of the semiconducting layer. Such a component can be produced in a technically simple manner. The source electrode and the drain electrode can be produced on a surface of the semiconducting layer by deposition.
The source electrode and the drain electrode may be at opposite ends of the semiconducting layer. In this way, a suitable distance can be provided between the source electrode and the drain electrode.
To produce the component in a technically simple manner, the component may comprise a substrate. The semiconducting layer is then applied above the substrate. The semiconducting layer may be a thin layer. The substrate may consist of silicon. There may be an electrically insulating layer between the substrate and the semiconductor. Interfering influences may be prevented by the electrically insulating layer.
Ferroelectric material of the component may be in the form of a layer. A layer of ferroelectric material may therefore be applied on the source electrode. A layer of ferroelectric material may therefore be applied on the drain electrode. A ferroelectric layer may be partially located on the semiconducting layer to enable technically simple production. A ferroelectric layer may be partially located on the source electrode and/or drain electrode. There need not necessarily be two spatially separated layers of ferroelectric material. It can be a single layer of ferroelectric material, which is at least partially on the source electrode and at least partially on the drain electrode. This can also facilitate producing.
A first and/or second gate electrode may be applied on the opposite side of such a ferroelectric layer. The first gate electrode and/or the second gate electrode are located on one side of the ferroelectric layer. The source electrode and/or the drain electrode are located on the opposite side of the ferroelectric layer.
A drain electrode consisting of metal may serve as substrate of the synaptic component. Semiconductor material may first be applied on the substrate. Above the semiconductor material, the metal for a source electrode may be applied. Ferroelectric material may be applied on the source electrode. Metal for a gate electrode may be applied on the ferroelectric material. A vertical construction is thus possible.
One or more electrically insulating layers may be provided during producing to produce a synaptic component in a technically simple manner. For example, an insulating layer with recesses may be applied on a semiconducting layer. The semiconducting layer may be contacted through the recesses. The recesses may be at least partially coated with metal to produce Schottky diodes, for example. The insulating layer with the recesses helps to keep the two electrodes electrically separated. Producing one or more insulating layers may also be advantageous in a vertical construction. The electrically insulating layer may cover the semi-material, as far as the semiconducting material is not to be contacted by metallic electrodes.
A plurality of synaptic components can be produced in a simple manner, which are electrically connected in parallel and/or in series. A so-called crossbar structure can be produced, which can be advantageous in neural networks.
The semiconductor material may be selected from: Si, Ge, SiGe, SiGeSn, GeSn, SiC. The semiconductor material may be a III-V compound semiconductor, II-VI compound semiconductor. The semiconductor material may be a 2D material, which thus consists of only one layer or only a few layers of atoms and molecules. The semiconductor material may be a substrate or a semiconductor layer on a substrate, such as a “silicon-on-insulator” (SOI). By substrate is meant a self-supporting layer, which can therefore serve as a carrier for other layers. For producing the component, it can therefore be assumed that there is a substrate on which further layers can be deposited. The semiconductor material may be a semiconductor heterostructure with several semiconductor layers.
The ferroelectric material may be selected from: HfO2-based ferroelectric, perovskite ferroelectric, and organic ferroelectric. The ferroelectric material, such as HfO2-based ferroelectric, may be doped. Doping may have occurred with one or more of the following elements: N, Al, Si, Sc, Ge, Y, Zr, Gd, Pr, Sr, Tb, La, Lu.
The metal for the Schottky diode may be selected from: Al, Ag, Au, Cu, Cr, Mo, Ni, Nb, Pt, Ti, Ni, TiN, TaN, and metal alloys. The metal for the Schottky diode may be a metal semiconductor alloy such as silicides, germanides, metal-SiGeSn alloys.
Through the invention, artificial synaptic elements based on metal/semiconductor Schottky barriers may be provided. The polarization of the ferroelectric layers leads to a shielding charge at the metal/semiconductor interface, so that the effective strength of the Schottky barrier in the metal/semiconductor contact is modulated. The polarization switching in multidomain systems induces multi-level charge distributions at the metal/semiconductor interface, causing a multilevel conduction of the diode. This can be exploited to imitate properties of synapses.
The present invention solves the problem of interface traps at the interface between ferroelectric and semiconducting material in FeFETs, because the ferroelectric material is primarily disposed on the metal. The ferroelectric devices according to the invention can be produced at relatively low temperatures (<800° C.) using the CMOS process. As a result, almost all ferroelectric materials can be used.
The invention provides a method of constructing artificial synapses, which may be produced according to the following pattern. They may comprise two metal/semiconductor Schottky diodes which are connected at the back and contacted with a ferroelectric layer and a gate electrode. The thickness of the ferroelectric layer on the two Schottky diodes may be identical or different. This also applies to the choice of material for the ferroelectric layer. One Schottky diode can act as a signal input, while the second Schottky diode can be used to control the synaptic weight. By applying a constant bias voltage to the second Schottky diode, the synaptic weight can also be adjusted via the control voltage on the first Schottky diode. Instead of a second Schottky diode, a metal-semiconductor junction may be present, which acts like an ohmic contact.
Additional effects occur with artificial synapses with reversed bias voltage of the metal/semiconductor Schottky diode, which are covered with a ferroelectric layer. Here the drain electrode can be realized by an ohmic contact at the semiconductor.
The invention also relates to a method for operating a synaptic component according to the invention. The first Schottky diode is connected in reverse direction (reverse biased) during operation. An electrical voltage is applied to the first gate electrode in a pulsed manner. The synaptic component then behaves like a biological synapse.
The figures show
S here is the integration over the input signals with the associated weights and 0 is an offset. The neural function ƒ(s) acts as a threshold function. Once S reaches a threshold value, the neuron with the function ƒ(s) is activated.
The metal of the source electrode 102 and the semiconductor material of the semiconducting layer 101 are selected such that the junction between the source electrode 102 and the semiconducting layer 101 is a Schottky contact. Thus, there is a first Schottky diode formed by the source electrode 102 and the semiconducting layer 101.
The metal of the drain electrode 103 and the semiconductor material of the semiconducting layer 101 may be selected such that the junction between the drain electrode 103 and the semiconducting layer 101 is a Schottky contact. Thus, the drain electrode 103 and the semiconducting layer 101 may form a second Schottky diode.
In an equivalent circuit, the first Schottky diode and the second Schottky diode are connected “back-to-back”. A potential difference between the electrical potential Vs and the electrical potential VD therefore results in one Schottky diode being connected in the reverse direction and the other Schottky diode being connected in the forward direction.
An electric current can flow from the source electrode 102 to the drain electrode 103 due to the potential difference between the electric potential Vs and the electric potential VD. When an electric current flows from the source electrode 102 to the drain electrode 103, the electric current passes the first Schottky diode in the reverse direction and the second Schottky diode in the forward direction. Flowing of the electric current in reverse direction can be done by electrons tunneling through the junction of the first Schottky diode. If the electric current flows through the second Schottky diode in the forward direction, the second Schottky diode acts like an ohmic resistor.
Above the source electrode 102 there is a first ferroelectric layer 104a. The first ferroelectric layer 104a may cover a part of the source electrode 102. The first ferroelectric layer 104a overlaps with the semiconducting layer 101.
Above the drain electrode 103A there is a second ferroelectric layer 104b. The second ferroelectric layer 104b covers a part of the source electrode 103. The second ferroelectric layer 104b overlaps with the semiconducting layer 101.
A first electrode 105a is located on the first ferroelectric layer 104a. A second electrode 105b is located on the second ferroelectric layer 104b. Both electrodes 105a and 105b may consist of metal.
The synapse is configured such that an electric potential Vg1 can be applied to the first electrode 105a. The Schottky barrier of the first Schottky diode may be modulated by an applied electrical potential Vg1. The synapse is configured such that an electrical potential Vg2 can be applied to the second electrode 105a. The Schottky barrier of the second Schottky diode may be modulated by an applied electrical potential Vg2.
The first electric potential Vg1 can be understood as a synaptic input signal within a neural network. The second electrical potential Vg2 may be used to control weights within a neural network.
The ferroelectric layers 104a and/or 104b may overlap regions of the electrodes 102 and 103, applied on the semiconducting layer 101. Production may thus be facilitated. Also, the Schottky barriers of the first and second Schottky diodes can be modulated in a more controlled manner.
In
Here ∅bn is the height of the Schottky barrier for the electrons. q is the charge of the electrons. ND is the doping concentration at the surface of the semiconductor 301.
If the voltage Vgi shown in
For a p-type doped semiconductor, the polarization and bias points in the opposite direction. The Schottky barrier is ∅bp·
The measurement curves shown in
The first Schottky diode 602, 601 is operated under a reverse bias voltage. This synaptic element operates like a single gate transistor, which can be represented by the symbol of
The structure shown in the
Number | Date | Country | Kind |
---|---|---|---|
10 2020 207 439.9 | Jun 2020 | DE | national |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/EP2021/065297 | 6/8/2021 | WO |