Integrated memristor ferroelectric memory developed for energy-efficient training of AI systems

Over the past few decades, electronics engineers have developed a wide range of memory devices that can safely and efficiently store increasing amounts of data. However, the different types of devices developed to date each have their own trade-offs that limit overall performance and limit possible applications.

Researchers from the Université Grenoble Alpes (CEA-Leti, CEA List), the Université de Bordeaux (CNRS), and the Université Paris-Saclay (CNRS) recently developed a new memory device that combines two complementary components, known as memristors and ferroelectric capacitors (FeCAPs), that are typically used separately. The integrated memristor ferroelectric memory, presented in a paper published in Nature Electronics, could be particularly promising for running artificial intelligence (AI) systems that learn autonomously to make more accurate predictions.

“The ‘ideal’ memory is one that is dense, nonvolatile, nondestructively readable, and has virtually infinite durability,” Elisa Vianello, senior author of the paper, told Tech Xplore.

“Unfortunately, such memories do not yet exist and may never be fully realized. We realized that addressing this limitation requires new approaches that can combine the complementary strengths of different memory devices.”

When investigating the components of existing memory solutions, Vianello and colleagues noticed that FeCAPs and memristors share a very similar stack structure, even though they rely on very different physical mechanisms to operate. This ultimately led to the development of a new memory that integrates the functionality of both these components within a single stack. This could be advantageous for implementing energy-efficient training and AI algorithms.

“Memristors store information by changing electrical resistance through the creation and dissolution of a conductive filament that connects two electrodes,” Vianello explained.

“Programming these resistance states requires precise current control, which affects both the programming power and the write endurance of the device. In contrast, read operations require only short pulses of low voltage to determine the stored resistance value.”

FeCAP, the second component of the hybrid system developed by Vianello et al., is a memory device based on ferroelectric materials. The device stores information through the reversible polarization of a ferroelectric material that is switched by an applied electric field.

“Because polarization reversal requires ultra-low displacement currents, FeCAP offers excellent durability and extremely low energy consumption during programming,” Vianello said.

“Our ferroelectric memristor memory integrates a silicon-doped HfO₂ (commonly used in FeCAP) and a titanium scavenging layer (typically used in memristors). In our design, all devices initially behave as FeCAPs, but thanks to the Ti layer, they can be converted into memristors through an electrical forming operation.”

Essentially, the hybrid memory created by this team of researchers combines the best features of memristors and FeCAP technology. Like memristors, they are good for inference because they can store analog weights, are energy efficient during read operations, and support in-memory computing. However, integrating FeCAP also supports fast and low-energy updates, making it ideal for training machine learning algorithms.

“We have demonstrated a memory technology that combines memristor and FeCAP functionality within a single stack,” Vianello said. “This hybrid approach leverages the strengths of both device types and enables efficient and reliable on-chip training and inference of artificial neural networks.”

This recent work by Vianello and colleagues may soon inspire other research groups to develop other hybrid data storage solutions that combine seemingly very different memory components. In the future, the memory they created could be further refined and used to support the training and implementation of edge AI, systems where AI algorithms run directly on local hardware, rather than relying on remote cloud servers or data centers.

“Many real-world applications require AI systems to continuously learn new tasks or adapt to changes in input without catastrophically forgetting previously acquired knowledge,” Vianello added.

“However, classical deep learning methods tend to overwrite existing parameters with new information. Recently, several new algorithms have been proposed to tackle these challenges. Our next goal is to integrate memory techniques with these new approaches, paving the way for systems that can continuously learn and dynamically adapt, much like the human brain.”

© 2025 Science X Network