Edge Intelligence: materials and devices

1. Edge Intelligence: materials and devices

The handling of a large number of inhomogeneous and unstructured data (text, video, sound etc.) requires a large computing power as well as an unconventional, intelligent computation and data storage approach subject to energy resources constraints.

Our target is to develop a technology platform based on ferroelectric Hf(Zr)O2 ranging from scalable non-volatile embedded memories to memristive electronic synapse and neuron devices. HfO2 is a key gate dielectric material that has enabled the scaling of nanoelectronics devices. The discovery of ferroelectricity in this material has opened the way for a number of other applications as briefly mentioned above taking advantage of the compatibility of this material with Si processing.

1.1 Energy efficient embedded ferroelectric non-volatile memories for IoT

The aim is to develop embedded Hf(Zr)O2 ferroelectric nonvolatile memories (NVM) for “normally off” Microcontroller Units (MCU) used in IoT sensor nodes and other edge applications. The new NVM can replace the currently used eFlash allowing a faster and energy efficient storage and retrieval of data. Our team, at ESSL/INN in the framework of H2020 project 3eFERRO-(GA No 780302) (site:, twitter: @3eFerro) investigates metal-ferroelectric-semiconductor (MFS) devices on Ge substrates with the aim to improve the endurance of memory FeFET devices. The Hf1-xZrxO2 (HZO) ferroelectric is obtained by plasma enhanced atomic oxygen deposition in the MBE system.  Symmetric hysteresis curves free of “wake-up” effects and record high remanent polarization has been achieved, attributed to the clean Ge/HZO interfaces and the beneficial influence of the substrate. Our plans for the next couple of years will be to fabricate and evaluate p-channel Ge FeFET targeting devices with improved reliability over their Si counterparts.

1.2 Ferroelectric electronic synapses for neuromorphic processors

It has been found that by voltage controlled partial polarization it is possible to obtain a multistate analog memory operation in ferroelectric field transistors (FeFETs) and ferroelectric tunnel junctions (FTJ) which make them suitable for memristive electronic synaptic devices used for on-line training of deep neural networks.  Our team at ESSL/INN, in the framework of the H2020 project BeFerroSynaptic-(GA No 871737) (site:, twitter: @BeFerroSynaptic) explores the ultimate  limits of (epitaxial) Hf(Zr)O2 ferroelectric thickness in order to make sensitive and ultralow power FTJ memristive devices. Our optimized devices are planned for integration in the back-end-of line of CMOS circuitry by our industrial partners to realize new ferrosynaptic neuromorphic processor designs.