MRS Fall Meeting 2020

Scaled VO2 Oscillators for Neural Network Applications

View publication


Oscillatory Neural Networks are a novel, promising concept as hardware accelerator for neuromorphic computing. Systems of coupled oscillators present associative memory capabilities that can be used to implement tasks such as image recognition [1]. We fabricated compact electronic relaxation oscillators using the phase change properties of VO2, which undergoes a temperature driven metal-to-insulator Mott transition. We studied and developed fabrication techniques for the VO2 coupled oscillators on a Si platform, to retain compatibility with CMOS process. VO2 films were deposited with Atomic Layer Deposition on a SiO2/Si substrate. The as-deposited amorphous VO2 films crystallize during a post-annealing treatment under oxygen flow at a temperature of 450° for 20 minutes. Typically, polycrystalline films with an average grain size between 80 and 50 nm [2] form on the SiO2 surface. The annealed films show a phase transition at 340 K with a resistivity change of 2 to 3 orders of magnitude and a hysteresis width of about 10 K. The polycrystalline film was patterned with Ebeam lithography and ICP etching to define scaled VO2 oscillator structures, with dimensions between 200 and 500 nm. The oscillators are tested in their switching characteristics, and the variability impact on their coupling dynamics is addressed. In particular, we investigated the phase transition of the polycrystalline devices using electrical characterization and Scanning Thermal Probe Microscopy (SThM) [3] on a single grain level. The temperature maps obtained with the SThM provide nanometer resolution and allow to observe the thermal signature of single grain switching in the scaled oscillator devices. Hence, the current path can be visualized in the insulating and metallic phases of the VO2 device. The microscopic phasechange properties could be correlated with the characteristic waveforms of the oscillating pattern in a coupled oscillator system. In addition, a milli-second flash-anneal technique was tested and optimized to control the crystallization process. The average grain size could be reduced from 55 nm to 20 nm, resulting in more uniform electrical device characteristics and allowing further scaling of the VO2 oscillators. The devices were tested in vacuum and coupled through external electrical components. A network of coupled oscillators locks in frequency and establishes programmable phase relations depending on the strength of the coupling [4,5]. We demonstrate the capability of the oscillator network coupled with an array of tunable resistors to perform image recognition. Experimental results of a three-coupled oscillator system and simulations on larger networks are shown illustrating the associative memory capabilities of the oscillating neural network. This work is supported by the HORIZON2020 PHASE-CHANGE SWITCH Project (Grant No. 737109). [1] F. Hoppensteadt and E. Izhikevich, Phys. Rv. Lett, vol. 82, 2983-2986, 1999. [2] A. P. Peter et al, Adv. Funct. Mater., 25: 679-686, 2015 [3] F. Menges et al., Rev. Sci. Instr., vol.87, no. 7, pp. 074902, 2016 [4] N. Shukla et al., J. Appl. Phys., vol. 117, no. 5, 2015. [5] E. Corti et al. 2018 IEEE International Conference on Rebooting Computing (ICRC), 2018, pp. 1-7. doi: 10.1109/ICRC.2018.8638626