Focus issue on hafnium oxide based neuromorphic devices


 Our pathway towards intelligent computing systems leads to an ever-increasing amount of data movement, which apparently pushes the conventional von-Neumann architecture towards its limits in terms of performance and energy consumption. Combining computing and storage functionality locally within one structure is seen as one potential branch to bypass this issue. Neural Networks are a very promising path in that direction. Artificial neural networks rely to a large extend on “synaptic weight” storage and multiply-accumulate (MAC) functionality in either digital or analogue way, the latter one making use of Kirchhoff’s and Ohm’s law. The brain inspired neuromorphic computing architectures go one step further and more directly mimic the key biological elements of the brain: neurons and synapses. For hardware realization of such neuromorphic computing architectures the availability of scalable non-volatile memory devices to realize high density synapses for deep learning artificial neural networks or brain inspired spiking neural networks is essential. This NCE Focus Issue concentrates on the discussion of hafnium oxide based neuromorphic devices. Hafnium oxide has become a standard dielectric material in complementary metal oxide semiconductor (CMOS) fabrication processes since its introduction as gate dielectric for metal oxide semiconductor field effect Transistors (MOSFETs) back in 2007. Since then, its possible application field has significantly widened into the usage as memory devices. It was shown that valence change based resistive switching devices, better known as either resistive random-access memory (RRAM) or Memristor, could be realized with good properties using hafnium oxide. Moreover, in 2011 it was reported that under special conditions hafnium oxide can even be transformed into a ferroelectric. The latter enables a variety of different types of memory cells, namely capacitor based ferroelectric random access memories (FeRAM), ferroelectric field effect transistors (FeFET) and ferroelectric tunneling junctions (FTJ). This special issue will cover all aspects of using hafnium oxide based devices in neuromorphic systems starting from the material optimization via device concepts and modeling towards simulation and integration of neuromorphic systems.

Vaccaro et al focus in their work on the physics-based modeling of the analogue switching dynamics of HfO x -based RRAM devices [3]. Their sophisticated approach allows to accurately describe the operative analog behavior of the filamentary resistive switches. The model succeeds both in replicating the dynamical analog and digital behavior and in highlighting all the main characteristics of the device such as abrupt SET and gradual RESET in quasi-static regimes, memory window in weak and strong pulse regimes, highly nonlinear kinetics with the correct slope over several orders of magnitude in switching times. Moreover, the model considers even the stochasticity involved in the filament formation and dissolution due to its relevant effect in dynamic conditions. Bridging device optimization and circuit design this model sets ground for detailed system simulations, thus supporting the crucial design-technology-co-optimization.
Wang et al discuss the insertion of an Al 2 O 3 layer into ultrathin bilayer layer stacks of HfO 2 /Al 2 O 3 in comparison to pure HfO 2 based structures [4]. They realize 1T-1R synapses. It turned out that a layer stack having 3 nm of HfO 2 in series to 1.5 nm of Al 2 O 3 showed the best properties and could outperform pure hafnium oxide stacks of 5 nm thickness. From their analysis of the conduction mechanism, they conclude that in the optimized bilayer stack a stable ionic conduction is present in contrast to the more complex conduction mechanism in the single layer stack. Finally, they characterize spike time dependent plasticity for both the optimized bilayer and the single layer stack and simulate the MNIST recognition accuracy for both cases. The system based on the bilayer stack shows a high recognition accuracy of 95.6%.
Finally, Brivio et al review structure, properties and applications of non-volatile as well as volatile HfO 2 -based RRAM devices in neuromorphic computing [5]. In this work, first the defect mediated filamentary switching mechanisms are discussed. Moreover, light is shed on the different programming algorithms for high-precision multilevel operation targeting at analog weight update in synaptic applications and for exploiting the resistance dynamics of volatile devices. Finally, neuromorphic applications such as artificial neural networks with supervised training and with multilevel, binary or stochastic weights as well as spiking neural networks are presented. From this overview and the results gathered in the other articles, HfO 2 -based RRAM appears as a mature technology for a broad range of neuromorphic computing systems.
Stepping apart from the filamentary RRAM devices, in the fifth article, Covi et al review the potential of ferroelectric HfO 2 -based devices-namely FeFETs and FTJs for their suitability in neuromorphic computing systems [6]. These devices typically feature a double-layer stack consisting of an about 10 nm thick ferroelectric HfO 2 -based switching layer and a dielectric layer that acts as an interface to the Si-channel in the FeFET devices, but is used as a tunneling layer in the bi-layer FTJ concept, thus allowing to decouple switching and tunneling layer. Gradual and analogue switching makes both device types interesting candidates for utilization in synaptic circuits whereas the accumulative switching characteristics are promising for the development of pulse accumulators in neuron circuits. Thanks to their non-volatility, CMOS compatibility, and extreme energy efficiency in a small footprint, these ferroelectric devices are a promising emerging technology especially for edge computing neuromorphic systems.
Bégon-Lours et al discuss a slightly different ferroelectric device architecture [7]. Ferroelectric synaptic weighting elements are fabricated using an ultra-thin (2.7 nm) hafnium-zirconium-oxide (HZO) layer. In these devices the ferroelectric switching and tunneling layer are synergized into one layer. This is made possible by utilizing a WO x -electrode. Moreover, crystallization of the HZO layer using millisecond flash-lamp annealing allows the direct integration of such devices into the BEOL of CMOS chips. In this work high current densities and resistive switching with voltage pulses as short as 20 ns at about ±1 V are demonstrated together with a small device-to-device variation as well as cycle-to-cycle variation below 1%, which silhouettes these devices from many competing concepts.
We hope that this article collection provides both, a comprehensive review of the versatile HfO 2 -based neuromorphic devices, as well as new insights into current research activities. Finally, the work in this exciting field might lead to the demonstration and application of neuromorphic architectures, advancing further on our pathway towards intelligent computing systems.

Data availability statement
No new data were created or analysed in this study.