Temperature effect on a weighted vortex spin-torque nano-oscillator for neuromorphic computing

Ren Li,Yasser Rezaeiyan,Tim Böhnert,Alejandro Schulman,Ricardo Ferreira,Hooman Farkhani,Farshad Moradi
DOI: https://doi.org/10.1038/s41598-024-60929-3
IF: 4.6
2024-05-04
Scientific Reports
Abstract:In this work, we present fabricated magnetic tunnel junctions (MTJs) that can serve as magnetic memories (MMs) or vortex spin-torque nano-oscillators (STNOs) depending on the device geometry. We explore the heating effect on the devices to study how the performance of a neuromorphic computing system (NCS) consisting of MMs and STNOs can be enhanced by temperature. We further applied a neural network for waveform classification applications. The resistance of MMs represents the synaptic weights of the NCS, while temperature acts as an extra degree of freedom in changing the weights and TMR, as their anti-parallel resistance is temperature sensitive, and parallel resistance is temperature independent. Given the advantage of using heat for such a network, we envision using a vertical-cavity surface-emitting laser (VCSEL) to selectively heat MMs and/or STNO when needed. We found that when heating MMs only, STNO only, or both MMs and STNO, from 25 to 75 °C, the output power of the STNO increases by 24.7%, 72%, and 92.3%, respectively. Our study shows that temperature can be used to improve the output power of neural networks, and we intend to pave the way for future implementation of a low-area and high-speed VCSEL-assisted spintronic NCS.
multidisciplinary sciences
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: How to improve the performance of spintronic neuromorphic computing systems (NCS) based on magnetic tunnel junctions (MTJs) through the temperature effect. Specifically, the author studied the influence of temperature on magnetic memories (MMs) and vortex - spin - torque nano - oscillators (STNOs) to explore whether temperature can be used as an additional degree of freedom to improve the output power of neural networks, and proposed a method of using vertical - cavity surface - emitting lasers (VCSELs) for selective heating to achieve a more efficient and rapid training process. ### Main research contents: 1. **Device fabrication and characteristics**: - Two magnetic tunnel junction devices with different geometries were fabricated, one for magnetic memory (MM) and the other for vortex - spin - torque nano - oscillators (STNO). - The resistance changes of these devices at different temperatures were studied, especially the resistance changes in the parallel and antiparallel states of MM. 2. **Influence of temperature on STNO**: - The output power and frequency characteristics of STNO at different temperatures under different bias currents were measured experimentally. - It was found that as the temperature increases, the output power of STNO increases significantly, but after a certain temperature, the output power begins to decline. 3. **Influence of temperature on MM**: - The resistance changes of MM at different temperatures were studied, and it was found that the resistance in the antiparallel state decreases with the increase of temperature, while the resistance in the parallel state changes less. - This temperature dependence can be used to adjust the synaptic weights in neural networks. 4. **Implementation and analysis of neural networks**: - A neural network composed of multiple MMs and one STNO was constructed, and the influence of temperature on the network output power was studied. - Through three different heating scenarios (heating only MM, heating only STNO, heating both MM and STNO simultaneously), the gain of output power and the energy consumption of laser heating were analyzed. 5. **Application examples**: - A neural network for waveform classification was simulated to show the influence of the temperature effect on the classification task. - The results show that through temperature regulation, the output power of the neural network and the signal - to - noise ratio (SNR) can be significantly improved. ### Main conclusions: - Temperature can significantly improve the performance of MTJ - based NCS, especially the output power of STNO. - Using VCSEL for selective heating can achieve efficient training and performance improvement without significantly increasing energy consumption. - Future research directions include improving device manufacturing processes, improving the energy conversion efficiency of STNO, and optimizing photonic components to reduce the energy consumption of VCSEL. Through these studies, the author provides new ideas and technical paths for future low - power, high - speed spintronic neuromorphic computing systems.