MemWerk - memristive materials for neuromorphic electronics

1. Outline

The steady progress of digitalization and the increasingly extensive use of artificial intelligence change our society, technologies and science in an entirely new way. However, this development is characterized by an ever increasing energy demand contributing significantly to the rise in CO2 emissions. Therefore, the energy consumption of our IT world turns out to be an ever more limiting factor of the Digital Revolution and an immense burden on our climate. Current technologies can hardly solve this issue and new innovative approaches are required. Here, a promising approach is neuromorphic electronics, in which biological learning and memory processes are reproduced electronically.  This approach is supported in the MemWerk project - memristive materials for neuromorphic electronics – funded for 5 years by the Carl Zeiss Foundation through the grant program "Durchbrüche" at the TU Ilmenau. In this interdisciplinary joint project, scientists from the fields of materials science, computer science, electrical engineering and information technology are researching how information processing in biological systems can be transformed into energy-efficient technical systems. The aim is to investigate the fundamentals of memristive materials for the tailor-made fabrication of memristive devices for bio-inspired (neuromorphic) electronics. Further information on this topic is available at

2. Motivation

The digital revolution (DR) and the widespread use of artificial intelligence (AI) are a changing trend in society, technology, and science in an unprecedented way. One of the major challenges from a scientific and technical point of view is the provision of a highly efficient IT hardware tailored to the special needs of DR and AI. Unfortunately, today's IT hardware concepts (Von Neumann computers based on Si-CMOS1) are far from optimal for many DR and AI applications and use significantly too much energy. Current projections show that in about 15 years the worldwide production of electrical energy would no longer be sufficient to cover the power requirements of the IT hardware then in use2. Bio-inspired neuromorphic systems for information processing based on materials with memristive properties are a promising approach for solving this problem. Therefore, the required materials are fundamentally different from those used in today's IT hardware and must be tailored to neuromorphic systems in order to get a much better energy efficiency.[1]

Our project starts at this point. The focus is on the material as a central component of information processing and storage (see Fig. 1). Memristive materials can change their electronic properties (e.g. electrical resistance) by external signals (e.g. current or voltage, light, temperature, etc.), whereby the change is maintained even after the signal has been removed. Hence, they provide a memory effect like synapses in neural networks. Memristive materials thus form the central structural component of artificial neural networks (KNNs), as shown in Fig. 1, and allow KNNs to be efficiently created in hardware.

[1]   Complementary metal-oxide-semiconductor. The (Si-) CMOS is generally understood to be the semiconductor process used to realize integrated digital and analog circuits.

[2]   Rebooting the IT Revolution: A Call to Action,

Figure1: Schematic representation of the interaction of an intelligent material (e.g. in an Artificial Neural Network KNN) with its environment. Here, it must adapt to different external stimuli.

The memory effect in memristive materials is achieved by microstructural changes in the atomic structure and ranges from the nano- and micro-scale to the macro-scale, whereby the mechanisms of change are based on multiple processes and their coupling. They include, for example, nano-structural changes, changes in defect structure, nano-ionic or solid-phase electrolytic mechanisms. The strength, frequency, and periodicity of the applied signals also play an important role for the resistance change. Thus, there is a unique potential for a detailed functional replication of cellular dynamics of biological learning processes, described as synaptic plasticity and providing the basis for memory formation in biological systems. However, due to the large number of coupled non-linear effects in memristive materials, a coherent physical understanding of material dynamics has not Complementary metal-oxide-semiconductor. The (Si-)CMOS is generally understood to be the semiconductor process used to realize integrated digital and analog circuitsComplementary metal-oxide-semiconductor. The (Si-)CMOS is generally understood to be the semiconductor process used to realize integrated digital and analog circuitsyet been achieved. In addition, the large number of technology and network parameters to be considered for the integration of memristive materials into neuromorphic systems increases the difficulty of targeted use of memristive materials in neuromorphic systems. Therefore, achieving the full potential of memristive materials for the neuromorphic hardware requires a holistic, interdisciplinary approach, setting the basic material, process, and technology parameters in direct relation to the desired functionality of the material within the neuromorphic system.

3. Objective

The objective of the project is to create the basis for a parameter-oriented development of memristive materials for use in neuromorphic systems. This should allow a target development of hardware-based cognitive electronics. For this purpose, a material atlas is to be developed through the digitalization of materials (i.e. the collection and evaluation of material, process and system data) as a novel "mapping system" of memristive materials, which relates material, process, and technology parameters directly to (i) the electronic properties of memristive devices (current-voltage characteristics, data retention, endurance, parameter variations, switching times, etc.) and (ii) the network characteristics (learning and memory performance of ANNs, i.e. learning and prediction tasks). The identification of key functions of materials and their response to external stimuli will permit the tailoring of material systems for neuromorphic systems. For this purpose, two particularly promising classes of materials, nano-ionic resistive switching materials and 2D materials, are to be worked on as examples. Accordingly, the materials must be designed, manufactured, characterised, and integrated into electronic devices and circuits.

Figure 2: Presentation of the work packages and their content networking along the two main topic axes.

4. Project activities

The activities within this project are structured around the objectives defined in chapter 3 (creation of a material atlas; tailor-made material systems for neuromorphic systems) and grouped in six work packages (WP) with the main focuses: material synthesis, technology, characterisation, systems, material atlas and modelling. These aspects are described in more detail in the following. For a better overview, the planned work is visualised in Figure 2.

WP1 Memristive materials: Within WP1, memristive materials are produced while varying various process parameters (e.g. O2 partial pressure in reactive sputtering processes, temperature, deposition geometries, deposition rates, deposition processes, etc.). The individual process and material parameters are gradually varied to cover the largest possible parameter space. For this purpose, e.g. material composition and layer thicknesses are varied over the wafer within one process run. To address the broadest possible functional range for the reproduction of plastic memory formation and to enable comprehensive parameter screening, the investigations focus on two disjunctive material classes: (i) 2D materials (with emphasis on TMDs) and (ii) memristive oxide materials based on transition metal oxides (e.g. HfOx, TiOx, AlOx).

WP2 Technological platform for memristive components: A technological platform for memristive materials is being developed using a multilayer technology based on 4'' wafers, converting them into functional devices. The memristive devices are integrated into neuromorphic systems by means of a tailor-made assembly and connection technology for a direct integration to the system.

WP3 Electronic characterization and material analysis: A comprehensive material screening is carried out within this work package. Therefore, the deposited materials from WP1 are analysed according to their material properties and the memristive devices produced in WP2 are measured automatically on wafer basis. The aim is to obtain well-founded statistics on materials and devices in terms of their material and electronic properties. The obtained characteristics are considered in a material modelling, which allows a detailed description of the functional material properties. Furthermore, parameterised device models are developed based on the data obtained here, facilitating a targeted design of neuromorphic circuits.

WP4 Memristive neuromorphic electronics: Important system requirements for the material and the devices are evaluated within this work package. These requirements are used for a target-oriented material and technology development in WP1 and WP2 as well as the provision of parameters for research and analysis functions in WP5. Furthermore, training algorithms (learning models) and network structures of neuromorphic systems are designed to meet the special properties of memristive materials.

WP5 Materials Atlas: The multitude of material microstructures and technology parameters or processes basically leads to a very large parameter space. In WP5, this parameter space will be systematically "mapped" by building up a comprehensive material and process knowledge base (material atlas) and accessed using research and data analysis methods. The activities are divided into two phases: In the first phase, materials, and processes (WP1, WP2) as well as memristive devices are first examined in concentrated parameter models. Network-based systems are the special subject of the second phase.

WP6 Modelling of memristive materials: WP6 aims at a description of memristive materials, covering their structural, electronic, and stochastic properties within the framework of simulation models. Furthermore, this work package supports the parameter identification by a data driven modelling of memristive materials. Thus, WP6 forms a bridge between WP3 and WP6 and simulatively accompanies the material development in WP1. In exchange with WP4, the simulation models of materials developed here are transferred into concentrated parameter models of devices.

5. Participating institutions

Department of Electrical Engineering and Information Technology (TU Ilmenau)
Electronics Technology
Univ.-Prof. Dr.-Ing. Jens Müller

Micro- and Nanoelectronic Systems
PD Dr.-Ing. habil. Frank Schwierz
Univ.-Prof. Dr. rer. nat. habil. Martin Ziegler

Dr.-Ing. Jörg Pezoldt

Advanced Electromagnetics (TET)
Univ.-Prof. Dr.-Ing. habil. Hannes Töpfer

Materials for Electrical Engineering and Electronics
Univ.-Prof. Dr. rer. nat. habil. Dr. h. c. Peter Schaaf   

Department of Computer Science and Automation (TU Ilmenau)
Databases & Information Systems
Prof. Dr.-Ing. habil. Kai-Uwe Sattler
Internet: https://

Institute of Micro- and Nanotechnologies MacroNano® (TU Ilmenau)
Interdepartmental Center of Micro- and Nanotechnologies

Managing Director: Univ.-Prof. Dr. rer. nat. habil. Stefan Sinzinger

6. Contact

Coordinator: Univ.-Prof. Dr. rer. nat. habil. Martin Ziegler
Technische Universität Ilmenau
Mikro- und nanoelektronische Systeme
PF 10 05 65
98684 Ilmenau

Phone: +49 3677 69-3711