PID2022-139586NB: Fabrication, characterization, modeling, and simulation of memristive devices and systems for neuromorphic computing (NeMeSys)
Duration: from 01 September 2023 to 31 August 2026 (36 months)
Abstract:
Artificial intelligence based on neuromorphic computing is arriving to almost every corner of our life. From smartphones to home appliances, from unmanned vehicles to face recognition, from financial data analysis to YouTube preferences, all electrical and computer
systems will include in the future a piece of technology able to evaluate a complex situation and take some kind of decision. This is currently performed using specialized processing units and high-level computer languages, but what if the core computations associated
with these tasks were performed by a single chip at much less cost and energy requirements? Why not take advantage of the lessons Nature teaches us and use a system that mimics the operations carried out by our brain?, that is, a system comprising artificial neurons and synapses able to establish connections and pass the information among them.
This objective can be reached using devices called memristors: i.e. resistors with memory. Devices that when connected forming an array are able to respond to a specific electrical stimulus according to how they were previously trained or more advanced, that can learn from new experiences or data. In this context, we have set up a coordinated project covering different technological areas.
First, one subproject (IMB-CNM) aims to obtain the most suitable technology to meet the specifications required in terms of yield and resistive switching performance of single memristors, complex configurations, and crossbar arrays. Beyond conventional devices based on high dielectric permittivity dielectrics, novel technologies such as inkjet printing or those based on flexible substrates will be also developed for transient electronics and wearable devices. The aging mechanisms and operating conditions affecting the reliability of the structures will be determined with the purpose of identifying the most favorable technological conditions.
A second subproject (UVa) will be in charge of investigating a wide range of devices and performing advanced electrical characterization. Novel dielectric films exhibiting ferroelectric and ferromagnetic behavior as well as mixture and nanolaminates will be explored. A number of techniques such as photocurrent measurements for bandgap determination, impedance spectroscopy for permittivity measurements, and spectroscopy of preexistent defects by thermal analysis will be used to characterize the fabricated structures.
A third subproject (UGr) will be devoted to physical and electronic simulation of both single and combined devices. This includes kinetic Monte Carlo, finite elements-based simulators for a wide variety of dielectrics and operating regimes as well as compact descriptions suitable for circuital treatment. Special emphasis will be put on statistical analysis of variability and noise issues.
Finally, a fourth subproject (UAB/UIB) will focus the attention on compact modeling of conventional and unconventional neural networks. Brain-inspired and quantum memristive systems will also be explored along with the inclusion of new training strategies for neural networks considering the physical peculiarities of memristors. To this end a complete simulation platform which combines the programming language Python with circuital simulation in Spice will be considered.