Revolutionizing Artificial Neural Networks: Self-Powered In-Sensor ANNs with Molecular Ferroelectric Photomemristors

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Jul 12 2023

In an article in the press with the journal npj Flexible Electronics, researchers demonstrated in situ hardware implementation of self-powered in-sensor artificial neural networks (ANNs) using molecular ferroelectric (MF)-based photomemristor arrays and fabricated reconfigurable artificial retina using MF-based arrays.

Background

The advent of advanced communication technology has significantly accelerated the development of the Internet of Things (IoT), in which billions of sensors/devices are involved in an extensive data network.

Computer vision/computers obtaining data from visual inputs and artificial intelligence (AI) were incorporated into IoT devices for automatic decision-making, data acquisition, and extraction of necessary information due to the generation of massive amounts of redundant data.

However, the conventional Von Neumann architecture, specifically ANNs, requires a significant amount of power and time for AI computing owing to the Von Neumann bottleneck/isolated processors and memories. Recently, memristor array-based ANN hardware with processing and memory capability was implemented for in-memory computing.

Additionally, visual neural network computing/intelligent computer vision was proposed by introducing the photosensing functionality into in-memory computing to realize bioinspired visual processing of information similar to human brains and eyes. Sensors can form an ANN by considering tunable photoresponsivity as ANN weight to synchronously process and sense images in real time. However, locally storing the weights in the sensor is highly challenging.

Moreover, frequent extraction of weights from the external memory leads to higher power consumption and additional latency. Multiple tunable photosensors for AI vision have been demonstrated in several studies. Different techniques were introduced in these in these photosensors to improve their programmability, power, and speed.

However, most of these studies characterized the photoresponse on a single device and constructed array-based networks using computer simulation in place of a hardware implementation for AI vision. To overcome these challenges in developing in-sensor ANNs, a new device must be developed that possesses scalability, a uniform array, and sufficient photoresponsivity while simulated by visible light.

The inorganic ferroelectric photovoltaic effect has been utilized successfully for local sensor weighting to develop sensors serving as neurons that integrate the functionalities of processing, memory, and sensing. However, toxicity, intrinsic wide bandgap of insulating ferroelectrics, and processing technique issues, specifically high cost, have limited the application of insulating ferroelectrics for in-sensor applications.

Lead-free and non-toxic MFs possessing high ferroelectricity can be a suitable alternative to insulating inorganic ferroelectrics. Several MFs can sense normal visible light owing to their semiconducting properties. Biocompatibility, flexibility, scalability, and affordability are the other major advantages of using MFs to develop in-sensor ANNs. Thus, these devices can address the challenge of hardware implementation of ANNs, specifically for organic-based ANNs.

In-situ hardware implementation of in-sensor ANNs

In this study, researchers used MF-based photomemristor arrays to implement the hardware of self-powered in-sensor ANNs for in-situ optical signal recognition and acquisition.

The device architecture of the interfacial type MF photomemristors was realized by thermally depositing organic semiconductor/copper phthalocyanine (CuPc) layers on different substrates with interdigitated electrode arrays that had a channel length of two μm.

Subsequently, MF diisopropylammonium bromide (DIPAB) films were blade coated on the CuPc films. Scanning electron microscopy (SEM), piezoresponse force microscopy (PFM), and X-ray diffraction were used for characterizing the synthesized samples.

A 5 × 5 photomemristor array was synthesized as a reconfigurable artificial retina with the tunable photoresponse and conductivity of a lateral semiconductor/MF heterojunction for optical letter image recognition. Researchers achieved the photoresponsive functionality for the illumination of white light by adding a molecular semiconductor layer on the MF layers.

Additionally, the sense-memory-computation functionalities, including the convolution sum of photocurrents from several devices, light sensing, and nonvolatile optoelectronic memorization, were integrated into a single device array. Tunable ferroelectric depolarization was introduced into the ANN to realize reconfigurable photoresponse and conductance.

Significance of the study

In this study, in-sensor ANNs with continuously reconfigurable photoresponsivity and conductance were fabricated successfully using CuPc/MF DIPAB interfacial photomemristors. ANN hardware was also demonstrated for artificial retinas using MF-based photomemristor arrays.

The MF-based in-sensor ANN performed analog convolutional computation by treating photoresponsivity as synaptic weight and conducted recognition and perception of white-light letter images with low processing energy consumption. In each device, the photoresponse was linear, multi-level, and in-situ with light intensity, which laid the foundation for convolutional computation of in-sensor ANNs.

Negative/positive synaptic weights were expressed logically by a bi-directional ferroelectric polarization degree, which reduced the complexity of the hardware by half for conventional in-sensor neural networks. Depolarization-driven photoresponse resulted in low latency and zero processing energy consumption for in-sensor ANNs.

Researchers also successfully simulated a large-scale electrical and optical hybrid ANN for the regression and recognition of hand-written traditional Chinese characters. The recognition of the hand-written Chinese digits by the large-scale array demonstrated its scalability, ability of low power processing, and potential for applications in MF-based in situ artificial retina.