At the Leibniz Institute for Photonic Technologies and the Friedrich Schiller University Jena, Heidemarie Krüger is working on a technology that stores and processes data directly where it is created. This approach mimics the way the human brain processes information. Known as neuromorphic, this method avoids complex and energy-intensive transfers between processor and memory.
Neuromorphic systems and chips are inspired by how neurons and synapses in the brain process information. They use similar principles such as parallel data processing, adaptive learning processes, and energy-efficient signal transmission. The goal is to create powerful and energy-efficient hardware, particularly useful in fields like Artificial Intelligence, machine learning, or sensor technology.
Heidemarie Krüger is advancing research in this area through her startup, Techifab. With her team, she is developing so-called memristor-based components, which are both energy-efficient and powerful. This technology is expected to be used in many areas in the future. Krüger mentions applications for self-driving cars or smart industrial plants as examples.
A memristor (Memory Resistor) is an electronic component that not only conducts electric current but also “remembers” its resistance depending on the current flow. This means a memristor retains its resistance even when the power is turned off, allowing them to be used as both storage and processing devices. The use of memristors is not entirely new. The US startup Knowm experimented with them in 2015. However, the memristors used by Techifab differ from those used by Knowm.
According to the Science Information Service, the neuromorphic chip developed by Heidemarie Krüger stands out due to its memristor technology compared to traditional memory technology. These components function like synapses in the human brain: they store and process information simultaneously. Unlike conventional computers, which constantly move data between processor and memory, this technology works directly on-site. This saves energy and enables fast, decentralized data analysis.
“Memristors can process more than just ‘0’ and ‘1’ – they also handle intermediate states,” Krüger explains. This ability allows flexible data processing and opens up new perspectives for algorithms that simulate neural networks. The range of applications extends from predictive maintenance of machines to real-time analyses for safety-critical applications like autonomous driving.
The origin of this technology lies in an unexpected observation from 2011. During a material analysis, Krüger’s team discovered a characteristic loop curve – a clear indication of a memristor’s behavior. This property of “remembering” previous computing operations inspired the team to develop artificial synapses from a combination of bismuth and iron oxide.
Based on this, the vision of a functional chip emerged. Thanks to funding from the Federal Agency for Disruptive Innovation amounting to several million euros, Krüger was able to advance the development. “Our artificial synapses handle complex calculations like matrix multiplications extremely efficiently,” reports Krüger. Such calculations are essential for training modern AI systems and image processing algorithms.
In collaboration with the Technical University Bergakademie Freiberg, Krüger’s team is already testing the technology in initial pilot projects under real conditions. It has been shown that the neuromorphic chip can precisely detect even the smallest changes and reliably predict wear patterns.
While classical processors require more and more transistors to handle the increasing amount of data, conventional chip designs are reaching physical and energy limits. Neuromorphic systems, which combine memory and processing units, not only reduce energy consumption but also open up new possibilities for AI applications. “Our goal is not only to analyze data but also to learn, recognize patterns, and respond flexibly to new situations – without a constant connection to external data centers,” emphasizes Krüger.
This technology could not only make data centers more energy-efficient in the future but also enable AI systems with lower resource consumption. The current prototype has 32 memristors. In the next development stage, the number is expected to increase to over 200 to model more complex neural networks and further advance autonomous systems.