Science - USA (2022-02-04)

GRAPHIC: KELLIE HOLOSKI/

SCIENCE

science.org SCIENCE

INSIGHTS | PERSPECTIVES

orbital configuration in the
nickel atoms to largely reduce
the electrical conductivity.
Upon applying single-shot elec-
tric pulses across the material,
Zhang et al. could redistribute
protons within the lattice. This
generates a multitude of elec-
tronic states based on the final
distribution and local concen-
tration of the charges. These
metastable states can then be
configured on demand to per-
form the functionalities of resis-
tors, memory capacitors, neu-
rons, and synapses—a first for
a single device (see the figure).
In comparison with oxide-
based devices that rely on the
migration of oxygen vacancies
through the nickelate lattice,
the proton-redistribution ap-
proach used by Zhang et al.
enables a larger charge modu-
lation at a faster time scale because of
the much smaller ionic radius of protons
as compared with those of oxygen atoms.
Moreover, the devices are formed by using
semiconductor foundry-compatible tech-
niques and on substrates compatible with
complementary metal-oxide semiconduc-
tor (CMOS) circuits, making this a prom-
ising “lab-to-fab” technology and immedi-
ately pertinent to the electronics industry.
All electronic materials have defects that
can create a barrage of metastable conduc-
tivity states. As a result, the value of this re-
search discipline lies in the ability to identify
and engineer functional states for specific
applications. Zhang et al. discovered in their
device many metastable configurations, each
having distinct electrical and chemical sig-
natures. The authors first explore the possi-
ble configurations using theoretical calcula-
tions and subsequently confirm them using
a combination of resistance and capacitance
measurements, Raman spectroscopy, and
scanning near-field optical microscopy.
The large pool of possible electronic
states available within a single device
would help to enable the implementa-
tion of reservoir computing frameworks
in hardware. Reservoir computing is a
computational framework inspired by a
specific type of neural network theory
that maps input signals into higher-di-
mensional computational spaces by using
the dynamics of a fixed, nonlinear system
known as a reservoir. The hydrogen-doped
NNO of Zhang et al. is a strong candidate
to be used as nodes for reservoir comput-
ing in hardware. Each node has two non-
linear components: the memristor—a de-
vice that combines functions of a memory

and a resistor—and the similarly named

memcapacitor. Together, the two compo-

nents represent an internal state and out-

perform theoretical reservoirs on several

classification tasks in terms of improved

accuracy and faster convergence.

A reconfigurable device may also help

realize grow-when-required (GWR) net-

works, which is an artificial neural network

in which the system can, as its name sug-

gests, grow when required. In more techni-

cal terms, this means that when the input to

the GWR network does not achieve a certain

threshold of activity, the system would au-

tomatically create a new node. Through the

capacity to grow on demand, the network

overcomes problems caused by resource de-

pletion—a common problem for static com-

putational networks. Similarly, the network

can also shrink its size if inactive nodes are

detected, saving operational costs. According

to theoretical models created by Zhang et

al., their GWR network has supremacy over

static networks by up to 250% accuracy for

incremental learning scenarios. The recon-

figurability offered by the device further ex-

pands the efficiency and reliabil-

ity for a GWR system, at least

in theory, and will help realize

more dynamic architectures for

continual learning.

The reconfigurable device

by Zhang et al. represents a

substantial advance by having

multiple neuronal and synap-

tic functionalities embedded

within a single device. This can

enable compact and energy-

efficient neuromorphic system

designs of reservoir computing

frameworks and dynamic neu-

ral networks. However, to bring

this vision to practical hard-

ware implementation, research-

ers still have to find answers to

many questions, such as how

to deal with the nonuniformity

of the devices, how to make the

device connect to or disconnect

from the neural network, how to

rearrange the connections when the device is

reconfigured from one function to another,

and how to determine the role of each device

and apply the correct voltage scheme on it.

The electrical circuits in use today are

designed with multiple passive compo-

nents such as resistors, capacitors, and

inductors and active devices such as tran-

sistors. With the discovery of memristors,

circuit designers now have an extra degree

of freedom ( 10 – 12 ) when designing power-

efficient, high-performance systems.

However, from a material implementation

perspective, the construction of these com-

ponents still requires complex assembly of

various conductive, semiconductive, and

insulating materials. The ability to imple-

ment almost all of these elements with a

single material platform can substantially

change electronics. Hence, such reconfigu-

rable electronic devices could have positive

implications beyond neuromorphic com-

puting and machine intelligence. j

REFERENCES AND NOTES

1. P. A. Merolla et al., Science 345 , 668 (2014).


2. M. Davies et al., IEEE Micro 38 , 82 (2018).


3. R. A. John et al., Nat. Commun. 11 , 1 (2020).


4. K. Janzakova et al., Nat. Commun. 12 , 6898 (2021).


5. G. Milano et al., Nat. Mater. 10.1038/s41563-021-
   01099-9 (2021).


6. J. Torrejon et al., Nature 547 , 428 (2017).


7. W. Yi et al., Nat. Commun. 9 , 4661 (2018).


8. R. A. John et al., Adv. Mater. 30 , 1800220 (2018).


9. H.-T. Zhang et al., Science 375 , 533 (2022).


10. S. Kumar et al., Nature 585 , 518 (2020).


11. S. Goswami et al., Nature 597 , 51 (2021).


12. Z. Wang et al., Nat. Commun. 9 , 2516 (2018).

ACKNOWLEDGMENTS

R.A.J. acknowledges the suport from the ETH Zurich

Postdoctoral Fellowship scheme.

10.1126/science.abn6196

Capacitor

Artificial

synapse

Resistor

Artificial

neuron

H+

“This can enable compact

and energy-efficient

neuromorphic system designs

of reservoir computing

frameworks and dynamic

neural networks.”

Hydrogen-doped perovskite nickelate as a

versatile reconfigurable platform

By applying electric pulses, the hydrogen ions in the nickelate lattice can occupy

metastable states and enable distinct functionalities. This allows the same device

to be reconfigured on demand as a resistor, a memory capacitor, an artificial

neuron, or an artificial synapse.

496 4 FEBRUARY 2022 • VOL 375 ISSUE 6580