(51) International Patent Classification 4 (11) International Publication Number: WO 89/ 012

H01C 10/00, H01G 9/22 Al G06F 15/46 (43) International Publication Date: 9 February 1989 (09.02.

(21) International Application Number: PCT/US88/02545 (81) Designated States: AT (European patent), AU, BE ( ropean patent), CH (European patent), DE (Eu

(22) International Filing Date: 27 July 1988 (27.07.88) pean patent), FR (European patent), GB (Europ patent), IT (European patent), JP, LU (European tent), NL (European patent), SE (European paten

(31) Priority Application Number: 078,836

(32) Priority Date: 28 July 1987 (28.07.87) Published

With international search report.

(33) Priority Country: US Before the expiration of the time limit for amending claims and to be republished in the event of the rec of amendments.

(71) Applicant: MAXDEM, INC. [US/US]; 267 South Fair

Oaks Avenue, Pasadena, CA 91105 (US).

(72) Inventor: MARROCCO, Matthew, L., Ill ; 2418 North

French Street, Santa Ana, CA 93706 (US).

(74) Agents: JOHNSTON, R., William et al.; Christie, Park¬ er & Hale, P.O. Box 7068, Pasadena, CA 91109-7068 (US).

(54) Title: ELECTRICALLY SETTABLE RESISTANCE DEVICE

(57) Abstract

A device (1) provides an electrically settable resistance, and comprises first (3) and second (5) electrodes. Each ele trode (3, 5) comprises a redox active material. The device (1) further comprises means (6, 7) defining an environment f each electrode (3, 5) such that the average redox state of the electrodes (3, 5), taken, remains substantially constant. The e vironment defining means (6, 7) includes an electrolyte (7) between the electrodes (3, 5). The device (1) further comprises pair of terminals (9, 11) electrically connected to the first electrode (3) at spaced-apart locations, the electrically settable r sistance being defined between the pair of terminals (9, 11), and being a function of the oxidation state of the first ele trode (3). The device (1) further comprises means, including a terminal (13, 15), electrically connected to the second ele trode (5), for defining a control circuit path, including the first (3) and second (5) electrodes and the electrolyte (7), t enable a control current flowing through the control circuit path to change the oxidation state of the first electrode (3) s that the control current flowing in one direction sets the resistance to a lower value and control current flowing in the o posite direction sets the resistance to a higher value.

ELECTRICALLY SETTABLE RESISTANCE DEVICE

BACKGROUND OF THE INVENTION

In general, this invention relates to an electrically settable resistance device; more particularly, it relates to such a device that is adapted to be micro-miniaturized and to circuit arrangements including a multitude of such devices.

^" _^ Circuit devices that exhibit memory have a wide and continually increasing spectrum of uses in . any kinds of circuit arrangements for performing various functions in suchdiverseapplications as communications, control systems, and computers. In such applications, the kinds of circuit arrangements in which memory devices are used include sequential switching circuits, in which the memory devices are interconnected with combinatorial logic circuits such as AND gates, OR gates, and the like. Other kinds of circuit arrangements, especially in the digital computer field, include registers within data processors and addres¬ sable memories in which many memory devices are intercon- nected in a matrix arrangement to provide for storage of - data.

For sequential switching circuit applications and for registers within data processors, the memory devices must be erasable, i.e., be adapted to change state repeatedly during circuit operation. Likewise, memory devices for a

random access memory (RAM) , need to be erasable so that new data can be written into the memory during data process¬ ing operations. Another type of memory, called a read-only memory (ROM) is written into before the memory is connected into the computer or other system, and has the limited function of providing read out of predetermined stored data. Highly desirable characteristics for a memory device include small size, low-power dissipation, long life, and high reliability. Substantial research and development has been directed for many years to make progress in achieving such desireable characteristics. Such research and develop¬ ment has involved diverse areas of technology, including magnetics, particularly with respect to magnetic cores on a large scale for a number of years and on an ongoing basis with moving magnetic media such as tape or disk; and including solid state physics, particularly with respect to junction transistors and metal oxide semiconductor transistors.

As to such progress, the integrated semiconductor chips used today contain many more memory devices than could ever have been practical using magnetic cores and generally operate faster and with less power dissipation. However, moving media storage such as magnetic tape and disk, and optical disk, is still generally required for mass storage. Commerciallyavailable "hard" or "Winchester" disks generally provide hundreds of times more storage capacitythan semicon¬ ductor memories. For example, there are personal computers presently on the market that accept an optional hard disk capable of storing in excess of 100 MB (Mega Bytes) . in comparison, a one megabit DRAM (Dynamic Random Access Memory) chip, eight of which are needed at a minimum for each byte, provides the maximum capacity for semiconductor memory currently available commercially. It has been reported that a four megabit DRAM chip has been made. As to a static RAM memory chip, its organization is such that

more transistors are required to define a memory cell. Because of its organization, such a static RAM chip generally has four- times less density than a DRAM chip.

Solid state memory has inherent advantages of speed, low power consumption, and high reliability over moving media memory such as tape and disk. In many applications, it is impossible or impractical to use any moving parts.

■There,-_are _. .many.:military and.θther _. applications, including deep space missions, that impose stringent requirements narrowing available options to solid state memory. . These stringent requirements include low power consumption, not only on average so as not to drain batteries or other sources of electric .power but also on a transient basis so as not to generate undue heat; fast access; radiation hard- ness; and high reliability.

Recent progress in increasing density in semiconductor memory is attributable to successes in reducing the size of the component transistors. It appears likely that further progress toward higher density in semiconductor memory by following this approach will become increasingly difficult. The smallest feature of these component transis¬ tors is now typically 1.5 microns or less. The ultimate limit of the optical lithography techniques in current use is about .0.5 micron. Smaller dimensions will require development of new lithographic techniques that have been suggested, such as e-beam or x-ray lithography. Furthermore, physical limitations place an ultimate limit on minimum transistor size. For example, the commonly used dopant concentration of 10 ¹⁶ cm" ^*3 provides only 10 dopant atoms in a cube 0.1 microns on a side. In addition, dealing with such factors as line resistance and edge capacitance becomes increasingly difficult as transistor size is reduced. As stated above, memory devices are incorporated within data processors as well as within random access memories. As incorporated within data processors such as

integrated circuit microprocessors, the memory devices are arranged into registers for storing program code and data manipulated in accord with the sequentially executed program code to produce an algorithmic solution to a problem. Digital computers, as conventionally organized, can by virtue of large memory capacity and high speed components compute solutions to arithmetic problems and retrieve -records from data bases far faster than is humanly possible. A digital computer program-includes written definitions of data structures and written program code for instructing a digital computer to perform sequential tasks in manipulat¬ ing data organized in accord with such defined data struc¬ tures. With respect to many problems, a person having ordinary skill in the art of writing digital computer programs can relatively easily and quickly write a digital computer program that a digital computer can execute to solve the problem. Particularly for applications such as accounting systems, word processing systems, and other systems that perform repetitious tasks, the amount of time it takes to write such a digital computer program is small compared to the amount of time saved by having the digital computer manipulate the varying data in accord with the same defined algorithms.

However, many problems within a broaά class of problems are impractical to solve by the conventional approach of writing and executing such digital computer programs. A representative type of suchproblem involves patternrecogni¬ tion. Even a young child can recognize a pattern far more quickly than even the most ^• powerful digital computer can with extremely elaborate programming. The ability of the human brain to solve such problems which are essentially impractical for a programmed digital computer to solve has led researchers such as J. J. Hopfield to analyze how signals propagate within nerve cells, and to propose con- struction of apparatus that would operate in an analogous

way. Among these proposals are various proposed types of networks that within a general class called an electronic neural network (ENN) . The literature concerning the various types of an ENN includes : J. J. Hopfield, Proc. National Acad. Sci.. Vol. 79, page 2554 et seq. (1982) ; and L. D. Jackel et al., J. Vac. Sci. Technol. B, Vol. 4, page 61 et seq.. (1986) ,* J". - J. -Hopfield- et al., Science. Vol. -233, .-page .625 et. seq. ;.;a-paper...by..T..J. Sejnowski et al., titled "Boltzmann -Machines: . Constraint- Satisfaction Networks That Learn," published by the Department of Computer Science, Carnegie Mellon University, as Technical Report CMU-CS-84-119 (1984) ; and a book titled "Parallel Processing, Volume 1 Foundations," published by MIT Press, Cambridge Mass. , 1986. Every ENN comprises a set of amplifiers interconnected through a network of ^~ ".resistors. The amplifiers may have various input/output ^" specifications; the output of an amplifier may be a linear or a non-linear function of its input. One type of ENN is called a Hopfield ENN, which has the following characteristics: 1) The amplifiers are non¬ linear; 2) In the general case, the output of every amplifier is connected through resistive elements to the inputs of all the amplifiers; 3) The resistive element connected between the output of any amplifier i and the input of any amplifier j has substantially the same resistance as the resistive element connected between the output of amplifier j and the input of amplifier i. In a Hopfield ENN, the input/output specification for each amplifier is usually such that the output is a sigmoid monotonic function of the input. Such a sigmoidal input/output specification is shown in the above-cited article in Science; see Fig. 3(A) at page 628. The curve shown in that Fig. 3(A) may be displaced along either axis; it also may be inverted about either axis. J. J. Hopfield has shown that a Hopfield ENN

may be described mathematically, and has certain desirable features. One such desirable feature is that, for any input state, there is a unique stable output state. The resistive elements used in a Hopfield ENN are adjusted to "program" the ENN for a specific problem, or to store information in the ENN for later recall.

In a Hopfield ENN comprising a set of N amplifiers, an array of N ² resistors must be provided. Each resistor must be settable to provide a -particular resistance value that corresponds to a strength of connection from one nerve to another. This requirement that the resistor be "settable" to a value means that the resistor must exhibit memory. The computing power of a Hopfield ENN increases with the number of resistors in the array. Accordingly, high density is highly desireable in this context,, just as high .density is highly desireable in a RAM. A Hopfield ENN with an array of 1024x1024 resistors becomes feasible only if the individual resistors can be made as small as about a 1 micron square. In addition to a Hopfield ENN, other types of an ENN are known. Another example of a type of ENN is called a layered neural network. A layered neural network typically has two or more sets or layers of amplifiers arranged in a hierarchy such that the output of each amplifier is connected to the inputs of only those amplifiers in a succeeding layer. Some of these amplifiers, especially those in the first layer, may be linear amplifiers. Such a layered neural network is a "feed forward" ENN because there is no connection from the output of an amplifier in a higher layer to the input of an amplifier in a lower" layer. An example of a layered neural network is described in, and referred to as a Boltzmann Machine, in the above-cited paper by Sejnowski et al. A layered neural network is often used as a learning network. This means that the resistive elements are iteratively altered during "training"

cycles in which the output derived from a known input is compared to a "desired output"; then an algorithm is used to compute new values for the resistive elements for the next iteration. .- With respect to the programming of an ENN, the Jackel et al. reference cited above discloses an ENN in which each resistor has an individually fixed value to define pre-programmed synapses _* ; ^■ .-._Other ENN's. that .are program¬ mable have been built using numerous switching transistors for selectively connecting and disconnecting resistors into and out of the neural net to provide programmable binary synapses; however, each such binary synapse has required 10 to 20 transistors. ..

It is desirable to provide not only programmable synapses but also more than two resistance values per synapse as is the case with a binary synapse. Providing for more than two resistance values per synapse would be very difficult with existing technology. . Extending the switchingtransistorapproachtaken inprovidingprogrammable binary synapses would entail hundreds of transistors per synapse, and this would severely limit the number of synapses for the neural net.

SUMMARY OF THE INVENTION This invention provides a novel and highly advantageous electrically settable resistance device and,circuit ^" arrange¬ ments including a multitude of such devices. The device is adapted to be micro-miniaturized and to be incorporated into any of a variety of circuit arrangements, and has particular advantages as a memory device for a high-density random access memory and as a settable resistor for a high-density resistor network.

The invention can be defined in various terms. Accord¬ ing to one definition, the invention resides in a device for providing an electrically settable resistance. The

device comprises first and second electrodes, eachcomprising redox active material. Means define an environment for each electrode such that its redox active material is reversible. The environment defining means includes an electrolyte between the electrodes. A pair of terminals are electrically connected to the first electrode at spaced- apart locations. The electrically settable resistance is defined between - the pair of terminals, and is a function of the oxidation state of the- first electrode. The device further includes means, including a terminal electrically connected to the second electrode, for defining a control circuit path, including the first and second electrodes and the electrolyte, to enable a control current flowing through the control circuit path to change the oxidation state of the first electrode so that control current flowing in one direction sets the resistance to a lower value and control current flowing in the opposite direction sets the resistance to a higher value.

As to the terminology used in the foregoing definition of the invention, "redox active" means that the material has at least two oxidation states which are accessible by either chemical or electrochemical means. Adevice embodying this invention may be used as a two-state device; i.e., as a switch that is electrically settable to provide either a high resistance or a low resistance. Alternatively, a device embodying the invention and replicated to form a resistance network of a neural net may be used to define a multiplicity of resistance values, with each of the many resistance values corresponding to a different one of many intermediate oxidation- states. Whether the material has two or more states, any two of the states constitute a redox couple. The term "reversible," further defines the redox active material as being capable in its environment, following oxidation or reduction between states of a redox couple, to re-reduce or re-oxidize essentially fully.

The reversibility of a given redox active material depends upon its environment, i.e., any contacting solid, liquid or vapor phase(s) , pressure and temperature, state of illumination or presence of other fields or gradients. For example, the material Prussian Blue/Berlin Green is a redox active material that can be.used for an electrode in a -device- embodying' this .invention ^' if the electrode is in an .appropriate ^τ environment in which, platinum is used as the material ^' terminals contacting' the electrode, and in which aqueous KCl is used as the material for the electrolyte in contact with the electrode. On the other hand, the Prussian Blue/Berlin Green redox couple is irreversible on platinum if aqueous tetramethylammonium chloride is the electrolyte. Further with respect to the term "reversible," and the converse term "irreversibility" as applied to a redox active material, each embraces both chemical and electro-- chemical matters. A redox active material is chemically irreversible if during oxidation or reduction a substantial portion of the material, such as a portion in excess of 1%, is converted into a reaction product that cannot be re-reduced or re-oxidized.

Electrochemical reversibility is a kinetic property, and can be quantified as the electrochemical rate constant for electron exchange, within the context of specified materials and environment. The time required to change the resistance setting of a device embodying this invention is an inverse function of the electrochemical rate constant, and is also a function of potential applied to produce the control current. It is undesirable for the applied poten¬ tial to substantially exceed the thermodynamic potential, for this results in energy inefficiency and production of waste.heat. A portion of the applied voltage is due to electrochemical irreversibility. That portion is called the Faradaic overpotential. A redox active material meets

the need for electrochemical reversibility for an electrode in an embodiment of this invention only if the time during which the resistance is to remain set is long compared to the time required to change the resistance setting, and if the Faradaiσ overpotential is less than one volt.

Each of a variety of redox active materials exhibit reversibility and is suitable for use in the electrodes of, and in the environment providedwithin a device embodying this invention. According to another definition, the invention resides in a thin film structure comprising a substrate supporting a multiplicity of cells, each cell providing an individually electrically settable resistance. Each cell includes at least three layers or deposits of materials. One deposit of material is a redox active material defining a first electrode. Another deposit of material is a redox active material defining a.second electrod . Another deposit of material defines an electrolyte between the first and second electrodes. The substrate further supports a multi- plicity of conductors arranged in a pattern to communicate electrically with a cell such that, at each cell two of the conductors are electrically connected to the first electrode at spaced-apart locations. These two conductors are referred to in the detailed description of a specific embodiment as a read (R) conductor and a common (C) conduc¬ tor. The electrically settable resistance of the' cell is defined between such two conductors. Further as to each cell, another of the conductors is electrically connected to the second electrode for defining a control - circuit path.- This conductor is referred to in a detailed descrip¬ tion of a specific embodiment as a write (W) conductor. The control current path includes the first and second electrodes and the electrolyte. A control current flowing through the control current path in one direction sets the resistance of the cell to a lower value, and in the opposite

directions sets it to a higher value. In accord with advantageous feature, a fourth layer or deposit is provided for each=cell. This fourth deposit has a non-linear resis¬ tance and electrically connects the W conductor to the second electrode so that in an operation to set the resis¬ tance of ^• . a' -cell _f - '-a,-potential ά_n; access =of a predetermined threshol ^" is"used"to 'set"a"cell resistance ^: on an-individual¬ ly, selected basis.

According to another definition of the invention, it resides in an electronic neural network comprising a set of amplifiers and multi-cell means embodied in thin film structure for inter-connecting amplifiers to establish the strengths of connection between the amplifiers. The multi- cell means defines a network of electrically settable resistances, each .of which is formed by a redox active material and has a resistance value that is a function of the oxidation state of the redox active material.

The foregoing and other advantageous features of the invention are described in detail below and recited in the appended claims. ■

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic illustration for the purpose of explaining some of the principles underlying this invention;

FIG. 2 shows in top plan view how a multiplicity of individually electrically settable resistance devices are incorporated in a thin film structure described in Examples

1 and 2 of the invention;

FIG. 3 is a semi-schematic plan view of the area indicated at 3 - 3* in FIG.2, that is enlarged to illustrate features of one of the individually electrically settable resistance devices;

FIG. 4 is a semi-schematic cross section taken along line 4 - 4* of FIG. 3, with some of the features being shown on an enlarged scale for clarity; FIG. 5 shows in a "top plan view how conductors and terminating pads are''laid down on a ^{^} substrate to form an array in the course of making a thin film structure described in Example 3 of the invention;

FIG. 6 is an enlarged plan view of an area, indicated at 6 - 6 - in FIG. 5, at a subsequent stage in the course of making the thin film structure;

FIG. 7 is an enlarged plan view of the area indicated at 7 - 7 - in FIG. 6; .

FIG. 8 is a cross section of the completed thin film structure, taken along line 8 - 8 ^« of FIG. 7, and shows structure within one of the individually, addressable and electrically settable- resistance devices;

FIG. 9 is a functional block diagram of the thin film structure of Example 3; FIG. 10 is a block and schematic diagram'of a model neural network for one type of electronic neural network embodying the invention; and

FIG. 11 is a block and schematic diagram of a layered neural network type of electronic neural network embodying the invention.

DETAILED DESCRIPTION

With reference to FIG. 1, a schematically illustrated multi-layer device 1 includes a first electrode 3 and a second electrode 5. Each of the electrodes 3 and 5 is made of a redox active material. Means including a covering layer. 6 and -an -electrolyte--7--define.--an environment for each" ^' electrode "-such"that its redox active material is reversible. ^' A:pair of terminals._9 and 11 are electrically connected to the first electrode 3 at spaced apart locations. An electrically settable resistance is defined between the terminals 9 and 11. This resistance has a value that is a function of the oxidation state of the reversible redox active material of the first electrode 3. Terminals 9 and 11 can be connected into external circuitry so that signal current can flow in series through terminals 9, the resis¬ tance defined by electrode 3, and terminal 11. Either one of terminals 9 and 11 can also be used to define a node of a control current flow path through device 1, as explained more fully below. Another pair of terminals 13 and 15 are electrically connected to device 1 at spaced-apart locations on electrode 5. Another electrically settable resistance is defined between terminals 13 and 15. This resistance has a value that is a function of the oxidation state of the reversible redox active material of the second elec- trode 5. Terminals 13 and 15 can be connected into external circuitry so that signal current can flow, in series through terminals 13, the resistance defined by electrode 5, and terminal 15. Either one of terminals 13 and 15 can-also be used to define a node of a control current flow path through device 1.

Device 1 provides a single cell redox battery in which the specific resistance of the materials of the opposite electrodes can be set. The resistance between terminals 9 and 11 is a function of the specific resistance of the redox active material of electrode 3. The resistance

between terminals 13 and 15 is a function of the specific resistance of the redox active material of electrode 5. Device _L ^~ has the advantage that it provides non-volatile memory; i.e., following the setting of a resistance to a desired value, the set resistancevalue canremain essential¬ ly unchanged without the need for any externally applied power.

To set a desired value for the resistance between terminals 9 and 11, a potential is applied across the layers of device 1. The external potential source can be connected across terminals 13 and 9, for example. A control current flows through device 1 in response to the applied potential in a direction that depends on the polarity of the applied potential. The control current can pass enough charge to fully reduce the redox active material of electrode 3. Being fully reduced, the redox material of electrode has a set specific resistance, and the resistance between terminals 9 and 11 has a corresponding set resistance value. Because the redox active materials used in accord with the principles of this invention are reversible, the setting of the resistance value can be repeatedly changed. To do so, the external potential source is used to apply a potential across the layers of device 1 with a polarity to cause control current to flow to re-όxidize the redox active material of electrode 3.

The contacting layers of device 1 provide a path through which electric control current can flow in response to a drive potential applied by an external source.

Various materials may be used in making multi-layer device 1. One redox active material that may be used in electrodes 3 and 5 is tungsten oxide (W0 ₃ ) , suitably deposit¬ ed as a one micron thick film. A suitable material for electrolyte 7 is polyethyleneoxide + lithium triflate (PEO+LiCF ₃ S0 ₃ ) , also suitably deposited as a one micron thick film. The terminals 9, 11, 13, and 15 are suitably

inert metal such as gold wires patterned by standard micro- lithographic techniques. Other electrically conductive materials that may be chosen for these terminals are electri¬ cally conducting oxides and ceramics such as doped tin oxide and high temperature superconducting materials such as YBa2C ₃ θ7. ^•

Another redox active material that may be used in electrodes 3 and :5.'.is -polyvinylferrocene, and another suitable material for electrolyte 7 is polyvinyl (meth- ylpyridinium chloride) .

The redox active material for electrodes 3 and 5 may also be chosen from the following: polypyrrole, poly-3-alkyl- thiophene (alkyl=C ₁ to C ₁₆ ) , polythiophene, polyaniline and substituted polyanilines, polyquinolines, polyquinazolines, polyacetylene and substituted polyacetylenes, polyphenylene and substituted polyphenylenes, polyphenylenesulfide, films and polymers, prussian blue and other metal cyanides, other metal oxides as oxides of titanium, zirconium, hafnium, vanadium, niobium, tantalum, chromium, molybdenum, tungsten, and manganese, polymers of ruthenium or osmium bipyridine/- vinylbipyridine complexes, Ag/Ag+ or any other material which can be redox cycled, and can also be made from dis¬ similar materials, or from combinations or materials.

The material for electrolyte 7 may also be chosen from other PEO+salt combinations, alkali metal b* 'alumina, nafion, g ₄ l5, and other ion conducting metal halides, and other fast ion conducting materials. The electrolytic material should have very low electronic conductivity if the resistance set must remain unchanged for any length of time.

With reference to FIGS. 2, 3 and 4, there will now be set forth two specific examples (Example 1 and Example 2) of the invention, in each of which a thin film structure 20 is made to provide a multiplicity of individually electrical- ly settable resistance devices. Thin film structure 20

could be incorporated in various circuit arrangements including but not limited to mass produced circuit arrange¬ ments for radio or TV receivers or the like which commonly incorporate factory set potentiometers to trim the circuit arrangement into specifications. Each individually electri¬ cally settable resistance device of thin film structure 20 is electrically insulated from the others, and can be used to replace a separate such potentiometer.

A thin film structure 20- made in accord with either Example 1 or Example 2 comprises a glass substrate 21 supporting twelve devices that each provide an individual¬ ly electrically settable resistance. Each device includes three conductors referred to as a write or "W" conductor, a read or "R" conductor, and a common or "C" conductor. Each conductor terminates at one of its ends at a pad such as pad 22 at one end of "a representative conductor 24. These pads are arranged about the periphery of the substrate, and provide for making solder connections. Each device further includes a first electrode 26, a second electrode 27, and an electrolyte 28. In each device, the R and C conductors define a pair of terminals electrically connected to first electrode 26 at spaced-apart locations, and the electrically settable resistance of the device is defined between these R and C conductors. ^" As to the W conductor of each device, it defines a terminal electrically connected to second electrode 27, and cooperates with the JZ conductor, the first and second electrodes, and the electrolyte to form a control circuit path for use in setting the resistance defined between the R and C conductors.

Example 1:

To make thin film structure 20, thirty-six gold conductors and thirty-six terminating gold pads are laid down on glass substrate 21 to form an array in the pattern shown in FIG. 2. In this pattern, there are twelve groups

of conductors, with each group comprising R, C, and W conductors. Each conductor is 100 microns wide and 2 microns thick. In the region where the R, C, and W conduc¬ tors are parallel to each other, they are on 200 micron centers.

^• The array and surrounding substrate are then coated with a '2.micron' ^" .thick ^•• film 'of 'a ^" redox active material, viz, poly-3-butylthiophene, by spin coating.' A layer of photo¬ resist is applied, patterned and developed by photolithog- raphy.. The areas not protected by the photoresist are then etched, either chemically or with reactive ion etching, to leave a pattern of islands of redox active material. Each, island either covers a portion of a W conductor and surrounding substrate, or covers and bridges betweenparallel portions of R and C conductors. As to the islands that cover a W conductor, each has a top surface that is general¬ ly rectangular, approximately 350 microns long (along the length .of the W conductor) , and approximately 150 microns wide (spanning across the 100 micron-wide conductor and 25 microns of surrounding substrate on each opposite side) . As to the islands that cover and bridge between R and C conductors, each has a top surface that is generally square, approximately 350 microns x 350 microns, with the bridging portion being approximately 100 microns x 350 microns x 2 microns. So configured, the bridging portion has an absolute resistance that is approximately 1400 ohms per ohm-cm of specific resistance of the redox active material. The islands of redox active material on the substrate are then partially oxidized either chemically or electrochemically. Then, a 5 micron thick film of an electrolyte, ^' viz, poly- vinylpyridinium chloride, is deposited over the center of the substrate. A layer of photoresist is applied, patterned and developed by photolithography. The areas not protected by the photoresist are then etched, either chemically or with reactive ion etching, to leave a pattern of isolated

cells. A final protective layer 29 of polyvinylidene chloride is then deposited. This produces 12 individually electrically settable resistance devices, each with three gold conductors leading to solder pads at the edge of the substrate. As to the environment for the redox active material of the first electrode, it is in contact with the R and C conductors (gold) , with substrate 21 (glass) , and with electrolyte 28 (polyvinylpyridinium chloride) . As- to the environment for the redox-active material of the second electrode, it is in contact with the W conductor (gold) , with substrate 21 (glass) , and with electrolyte 28 (polyvinylpyridinium chloride) .

The specific resistance of poly-3-butylthiophene varies by six orders of magnitude; in its oxidized form its specific resistance is 10~ ^2" ohm cm, and in its- reduced form its specific resistance is 10 ⁴ ohm cm. The reversible reaction it undergoes between its oxidized form and its

Oxidized Reduced

To establish a resistance setting for the individually electrically settable resistance device, a potential is temporarily applied across the W and C conductors to cause control current to flow in an appropriate direction to provide, the amount of charge needed to set the specific resistance of first electrode 26, upon which the absolute resistance between the R and C conductors depends. The maximum and minimum values to which the absolute resistance can be set, and the amount of charge needed for setting it, depend in part on the exact size and geometry of the device. For the device in the size and geometry specified above, in which a 2x350x100 micron film bridges between

the R and C conductors, the minimum value of the absolute resistance is approximately 14 ohms, the maximum value of the absolute resistance is approximately 14,000,000 ohms (14 Mohms) , and the maximum amount of charge needed to set the resistance (as when the resistance is being changed between the minimum value and the maximum value) is approximately

5Q miσrocoulombs..

As to reversibility, poly.-3-butylthiophene has good electrochemical reversibility-; its Faradaic overpotential is below 0.1 volt. Further, the environment defined by the materials in contact with the poly-3-butylthiophene elect¬ rodes is conducive to reversibility.

Example 2: _Λ - An array as described in Example 1 is prepared and coated with a film of redox active material, viz, tungsten oxide, by sputtering. A layer of resist is applied, patterned and developed to leave a pattern of redox active islands, as described above in Example 1. The islands of tungsten oxide on the substrate are then partially reduced either chemically or electrochemically to form tungsten bronze (Na _x W0 ₃ ) . Then, there are deposited, in sequence, a layer of sodium conducting glass, and a protective layer of silica. The specific resistance o.f Na _x W0 ₃ varies by over eight orders of magnitude. ^' As to reversibility, tungsten oxide has good electrochemical reversibility; further, the environment defined by the materials in contact with the tungsten bronze electrodes is conducive to reversibility. With reference to FIGS. 5, through 9, there will now be set forth another specific example (Example 3) of the invention in which a thin film structure 50 is made to provide 144 devices arranged in an array that is addressable for writing such that the resistance of each device can be electrically set on a individual basis. The 144 resistan-

S BST/TϋTH SHEET

of thin film structure 50 are interconnected to define a resistance network, rather than being electrically isolated from each other for independent reading. The features of this network are such that it has numerous applications for which non-volatile memory is desirable; one such applica¬ tion is within an ENN in which each resistance controls the strength of connection from the output of an amplifier ^" to the input of an amplifier. With reference to FIG. 9, the network within thin film'structure 50 defines twenty- four nodes; each of twelve C conductors Cl through C12 (not all shown), defines a respective one of twelve of these nodes, and each of twelve R conductors Rl through R12 (not all shown) defines a respective one of the other twelve nodes. An electrically settable resistance R]_ ^ defines ^" the strength of connection between conductor Cl and conductor Rl. The value of resistance R]_ ^ ^can ^ ^e preset to a desired value in advance of operation of the network- as part of an ENN. This presetting is accomplished by applying a potential, in excess of a predetermined threshold, across conductors Wl and Cl. In general terms, settable resistance i_j defines the strength of connection between conductor Ci and conductor Rj. Presetting of resistance ^ j is accomplished by applying a potential, in excess of a predetermined potential, across conductors Wj and Ci.

Example 3:

To make thin film structure 50, twenty-four gold con¬ ductors and twenty-four gold terminating pads are deposit- ed on glass substrate 51 to form an array in the pattern shown in FIG. 5. Each conductor is 10 microns wide by 1 micron thick. Twelve of these conductors are provided to serve as write conductors Wl through W12; the remaining twelve conductors are provided to serve as Read conductors Rl through R12. In one region of the substrate, the conduc-

tors are parallel to each other with the read and write conductors being interdigitated. In this region, the conductors are on 30 micron centers. This region is shown on an enlarged scale for clarity. The terminating pads for the write conductors are arranged in line with each other along..the edge.of substrate 51.-at the bottom of FIG. 5. The terminating pads for the read conductors are arranged in line .with, each.:other,lalong. the..-edge;of.substrate 51 at the top of FIG. ^' 5. "Then, ^" there is electropolymerized on every write conductor Wl through W12 a 1 micron layer of ruthenium bis-bipyridine-vinylbipyridine polymer. The ruthenium-containing polymer has a non-linear resistance, and is provided so that a threshold potential will need to be exceeded to perform a write operation to set an individual resistance value in use of the completed thin film structure. An additive material is silicon dioxide: a layer of silicon dioxide defines a non-linear resistance ^• by defining a tunnel -barrier. The read conductors Rl through R12 are not coated with the ruthenium containing polymer. Then, pyrrole is electropolymerize to cover all twenty-four conductors. Then, a layer of photoresist is applied, pat¬ terned and developed to leave a square array of 14 micron by 14 micron squares of electro-active polymer centered over the metal conductors. FIG. 6 shows a group of such squares on an enlarged scale. FIG. 7 shows, on a further enlarged scale, two such squares, one of which overlies write conductor Wj_ and the other of which overlies read conductor R^. Each square that overlies a write conductor has two layers, i.e., ruthenium-containing polymer layer 53 and polypyrrole layer 55. Each polypyrrole layer 55 is provided to serve as a redox active electrode electrically connected to the write conductor through layer 53. Each square that overlies a read conductor has a polypyrrole layer 57 that is provided to serve as a redox active elec- trode. At the stage of making thin film structure 50 to

which FIGS. 6 and 7 relate, each polypyrrole layer 57 is electrically connected to a read conductor; in the completed thin film structure each such layer is also electrically connected to a common conductor. Proceeding to the next stage, a layer of polyvinyl(methylpyridinium chloride)-co- polyvinylpyridine electrolyte is deposited, and patterned by photolithography such that each pair of electroactive polymer squares is bridged by electrolyte such as electrolyte 59 shown in FIG. 8. Then, insulating polyimide is applied and patterned to form a layer 60 from which depend 3 micron x 3 micron channels or vias 62, each of which is centered over and extends down to one of the polypyrrole electrodes for a read conductor. Finally, gold is deposited and patterned to form an array of twelve parallel gold con- ductors and twelve gold terminating pads, with the twelve gold conductors being- perpendicular to the first twenty- four conductors, such that contact is made, through the vias in the insulating polyamide, to the polypyrrole elec¬ trodes for the read conductors. These top conductors are the C conductors Cl through C12. Then, insulating polyamide is deposited to form a final protective layer 65, to complete the making of thin film structure 50. The 144 devices are individually addressable for writing using any pair of common and write lines. The resistance of each device is a function of the oxidation state of the electrode, such as electrode 57 that has a common conductor electrically connected to its top surface and a read conductor electrical¬ ly connected to its bottom surface. Each such resistance is adjustable over the range 10 ³ ohm to 10 ⁶ ohm. A fully charged device maintains its charge such that less than 1% of the charge is lost over a 24 hour period. Charging and discharging potentials are less than 150 millivolts.

Thin film structure 50 is a specific example in which the multiple cells define a square (12x12) matrix. Such a matrix can be made rectangular by, for example, providing

more pairs of write and read conductors than the number of common conductors.

With reference now to FIG. 10, a model neural network 100 includes a set of amplifiers A through A _n that are arranged in the same way as the amplifiers shown in the above-cited-article i -..Science , -b - .- r.J; -Hopfield, et al.- Each amplifier has an input, -a non-inverting output, and an., inverting .output«: ^: . Each non-inverting output produces a voltage signal referenced as ^ through V _n ; each inverting output produces a voltage signal referenced as vj through V^. Each amplifier has a non-linear input current to output voltage characteristic. A thin film structure 102 interconnects the amplifiers to establish strengths of connectionbetween the amplifiers. The specific interconnec- tion arrangement of model neural .network 100 conforms to a specific constraint. ^' That is, each amplifier is effectively connected to one and only one of the outputs of each of the other amplifiers. As to being "effectively connected" or not, this matter can turn on relative values of resistan- ces. A resistance that provides an effective connection has a substantially lower value than a resistance that does not. The common conductors C-_ through C _2n of thin film structure 102 are directly connected to the outputs of amplifiers A^ through A _n , and are connected to inputs I, through I _n . The read conductors RTL through R _n are connected as shown to satisfy the specific constraint set forth above. For clarity, FIG. 10 does not show the write conduc¬ tors that are used together with the common conductors in presetting the resistance values within thin film structure 102. Further, the resistances that are preset ^' to provide effective connections are symbolized in FIG. 10 by small squares at intersections of read and common conductors.

FIG. 11 shows in functional block diagram form an alternative embodiment in which an ENN 110 comprises a set of amplifiers that are interconnected by multi-cell means

1 to form a layered neural network. In FIG.11, the amplifiers are symbolized by circles, and the multi-cell means is symbolized by a group of blocks, of which block 112A and 112B represent sections of a thin film structure made in 5 accord with the same principles as the thin film structures described above. The pre-settable resistances of the thin film structure establish strengths of connections between the amplifiers.

For clarity, not all of- the amplifiers are shown in 0- FIG. 11, but rather only such representative amplifiers as provide for illustrating how the amplifiers are intercon¬ nected. As shown, an amplifier 114 in one layer of ENN ^' 110 is connected through individually pre-set resistances of section 112A to each amplifier such as amplifiers 116, 5 -118, and 120 at the next layer of ENN 110.

The number of I ^' ayers of ENN 110 may be chosen in accord with well known principles. Further, the number of amplifiers provided at any ^' given layer may be chosen in accord with such principles. For example, a layered neural 0 network type of ENN can be arranged so that 18 amplifiers

_ --are interconnected whereby 7 amplifiers in a first layer feeding forward to 4 amplifiers in a second layer, which .-- in turn feed forward. to 5 amplifiers in a third layer,

-.- which in turn feed forward to 2 layers in the fourth layer. 5 In such case, the resistances of the multi-cell means provide

28 effective connections from the first to the second layers, 20 more effective connections from the second to

-: the third layers, and 10 more effective connections from the third to the fourth layers. 0

5