Periodic Reporting for period 1 - FREEMIND (Ferroelectric REsistors as Emerging Materials for Innovative Neuromorphic Devices)
Reporting period: 2019-06-01 to 2021-05-31
Worldwide, data traffic is exploding. Internet of Things, commercial transactions, media… This is possible thanks to always more performant computers. However, the well-known “Moore’s law” governing this trend is now reaching a physical limit. It is now urgent to find new computing architectures. An efficient way to handle large datasets is to use Artificial Neural Networks algorithms, but the training of these networks on conventional computers is time and energy demanding. The overall objective of the research project “FREEMIND” is to develop a hardware dedicated to Artificial Neural Networks algorithms, functioning not in the digital but in the analog domain. More specifically, matrix-vector multiplications will be implemented in the shape of a cross-bar array of "memristors". Memristors are electrical components whose resistance can be changed in a non-volatile way by an stimulus, such as an electric field. Such components represent the "synapses" of the neural networks, in an analogy with the biologic synapses connecting neurons together in the brain. In "FREEMIND", the objective is to use a ferroelectric material to create a memristor. Highly strategic for IBM Research, the focus is on a CMOS compatible and scalable process.
The work performed during FREEMIND is interdisciplinary and required multiple techniques. Experimental research work includes experiment design, processing, characterization, and experimental set-up.
During the experiment design stage, programming skills were developed for the creation of parametrized lithography mask design. Programs were written in Python languages, allowing the generation of complex masks containing a broad variety of parametrized test structures, devices and crossbar arrays in a few hours of work time.
Processing work was at the heart of the project: on one hand, thin film deposition by Atomic Layer Deposition, optimization of annealing processes to obtain the ferroelectric phase of hafnium zirconate while remaining within CMOS-compatible thermal budgets, and materials characterization using X-Rays and microscopy. The crystallization of hafnium zirconate films as thin as 3 nm was demonstrated at only 450˚C. On the other hand, the development of device fabrication processes compatible with back-end-of-line integration, on a wide variety of materials stacks composed of nanometric thin films. Up to seven optical lithography steps were required for the fabrication of crossbar arrays of micrometric synapses. Using electrical characterization during the processing, the fabrication of 32x32 and 81x10 passive crossbar arrays without any shorted resistor was demonstrated.
Single devices were then electrically characterized on continuously evolving set-up, to measure the ferroelectric properties, and to demonstrate the synaptic behavior: the potentiation/depression was emulated by analog resistive switching obtained after applying voltage pulses of varying duration or amplitude. Temperature dependent transport experiments were performed in order to establish a model for electronic conduction through the memristors, providing guidelines for the optimization of the devices.
Finally, at the circuit level, crossbar arrays were characterized. The effect of crosstalk and sneak paths in passive crossbars were studied, and simple machine-learning experiments were performed.
Aside the experimental research work, literature studies covered the state-of-the-art on ferroelectricity in hafnium oxide, two and three terminals memristors in particular ferroelectric tunnel junctions, conduction mechanisms in oxides, analog computing… Learning on state-of-the-art in Neuromorphic Computing was enabled by the participation to international conferences.
During the experiment design stage, programming skills were developed for the creation of parametrized lithography mask design. Programs were written in Python languages, allowing the generation of complex masks containing a broad variety of parametrized test structures, devices and crossbar arrays in a few hours of work time.
Processing work was at the heart of the project: on one hand, thin film deposition by Atomic Layer Deposition, optimization of annealing processes to obtain the ferroelectric phase of hafnium zirconate while remaining within CMOS-compatible thermal budgets, and materials characterization using X-Rays and microscopy. The crystallization of hafnium zirconate films as thin as 3 nm was demonstrated at only 450˚C. On the other hand, the development of device fabrication processes compatible with back-end-of-line integration, on a wide variety of materials stacks composed of nanometric thin films. Up to seven optical lithography steps were required for the fabrication of crossbar arrays of micrometric synapses. Using electrical characterization during the processing, the fabrication of 32x32 and 81x10 passive crossbar arrays without any shorted resistor was demonstrated.
Single devices were then electrically characterized on continuously evolving set-up, to measure the ferroelectric properties, and to demonstrate the synaptic behavior: the potentiation/depression was emulated by analog resistive switching obtained after applying voltage pulses of varying duration or amplitude. Temperature dependent transport experiments were performed in order to establish a model for electronic conduction through the memristors, providing guidelines for the optimization of the devices.
Finally, at the circuit level, crossbar arrays were characterized. The effect of crosstalk and sneak paths in passive crossbars were studied, and simple machine-learning experiments were performed.
Aside the experimental research work, literature studies covered the state-of-the-art on ferroelectricity in hafnium oxide, two and three terminals memristors in particular ferroelectric tunnel junctions, conduction mechanisms in oxides, analog computing… Learning on state-of-the-art in Neuromorphic Computing was enabled by the participation to international conferences.
During the project, the demonstration of analog, non-volatile resistive switching was demonstrated in ferroelectric memristors fabricated with a back-end-of-line compatible process. The temperature during the process does not exceed 400C. The materials, Hf, W, Ti, Zr and O are present in the industry lines, and the processes are easily transferable for large-scale integration. There are two potential impacts: first, the synapses fabricated during FREEMIND can be cointegrated to CMOS neurons for the fabrication of neuromorphic chips. Second, the process, developed in the BRNC clean room, can be upscaled to industrial level.
In direct collaboration with a pre-doctoral researcher from ETH, Mattia Halter. Three-terminals ferroelectric memristors were demonstrated. For the first time, hafnium-zirconate based field-effect transistors were obtained with a Back-End-Of-Line, CMOS compatible process (Halter et al., ACS Appl. Mater. Interfaces 2020, 12, 17725−17732).
A second generation of FeFETs is currently being developed, with a different method for the fabrication of the metal-oxide channel. Larger memory window and retention time are obtained.
The ferroelectric origin of the analog resistive switching in a first generation of two-terminals devices, based on a TiOx interlayer, was demonstrated: a result published in (Begon-Lours et al., Physics Status Solidi, Rapid Research Letters 2020).
Finally, the combination of various experiments brings valuable light on the physical mechanisms explaining the resistive switching in the novel two-terminals devices proposed, and differs from what is commonly seen in ferroelectric tunnel junctions. A manuscript is currently in preparation for the dissemination of such results.
In direct collaboration with a pre-doctoral researcher from ETH, Mattia Halter. Three-terminals ferroelectric memristors were demonstrated. For the first time, hafnium-zirconate based field-effect transistors were obtained with a Back-End-Of-Line, CMOS compatible process (Halter et al., ACS Appl. Mater. Interfaces 2020, 12, 17725−17732).
A second generation of FeFETs is currently being developed, with a different method for the fabrication of the metal-oxide channel. Larger memory window and retention time are obtained.
The ferroelectric origin of the analog resistive switching in a first generation of two-terminals devices, based on a TiOx interlayer, was demonstrated: a result published in (Begon-Lours et al., Physics Status Solidi, Rapid Research Letters 2020).
Finally, the combination of various experiments brings valuable light on the physical mechanisms explaining the resistive switching in the novel two-terminals devices proposed, and differs from what is commonly seen in ferroelectric tunnel junctions. A manuscript is currently in preparation for the dissemination of such results.