Complex Optimization Problem Solving with Low Variability Neuromorphic VO2-Based Oscillators

Abstract

Recent advancements in phase-transition materials are paving the way for novel integrations in brain-inspired computing architectures.[1], [2] The pursuit of efficient, low-power neural processing systems mimicking biological brains’ fundamental operations is driving the exploration of novel 'neuromorphic' materials and devices. [3], [4] Among the most promising candidates, vanadium dioxide (VO2) stands out due to its unique properties. VO2's polycrystalline morphology, compatible with CMOS technology, enables it to switch between resistive states when electrically activated at low power through Joule heating, triggered when the internal temperature reaches 68 ˚C. [5] Self-sustained oscillations in VO2 devices provide means to design interconnected electronic oscillators with tremendous potential for neural network architectures, AI applications, and optimization tasks. [1], [3], [6], [7] Oscillation-based computing is highly effective for solving complex optimization problems (COPs), which typically require extensive computing resources, long processing times, along with even larger energy consumption figures. [8], [9] Our in-house fabricated VO2 devices have been used to design complete oscillating neural networks (ONNs), which we leverage to compute COP solutions within just a few oscillation cycles (< 25) and reduced power consumption. [6], [10] Specifically, the networks have been successfully applied to solve Graph Coloring problems and have embodied Ising machine solvers to compute Max-cut and Maximum-3Satisfiability (Max-3SAT) problems by implementing a novelly improved frequency-harmonic injection locking technique (SHIL). [6], [11] Additionally, we have experimentally demonstrated how the phase-transition characteristics in VO2 crossbar devices can be integrated with commercially available electronic components to create a biologically realistic FitzHugh-Nagumo oscillator. This oscillator generates complex waveforms similar to those in neurons, providing superior performance for computational tasks such as pattern retrieval and offering new possibilities for addressing and modeling complex time-dependent dynamic computing problems. [12] To achieve these results, we developed an optimized fabrication process at IBM Research Europe – Zurich, which allows precise control over the formation of granular nanostructures and the switching behavior of individual grains. [10] The devices were characterized using Raman spectroscopy, Atomic Force Microscopy (AFM), X-Ray Reflectivity (XRR), and current-voltage (I-V) and resistance-temperature (R-T) measurements. These capabilities enable the production of devices tailored for specific applications, which can be easily co-integrated with CMOS peripheral circuitry. Our VO2-based oscillators excel in local data processing, thereby avoiding the von Neumann bottleneck arising from energy-intensive data transfer between memory units and processors. [2], [4] VO2-based oscillator networks present compelling and scalable computing units for hardware accelerators, thanks to their high-performance switching properties and compatibility with Silicon. This project has received funding from the EU’s Horizon 2020 program under project No 101092096 (PHASTRAC). References [1] S. Carapezzi et al., “Advanced Design Methods From Materials and Devices to Circuits for Brain-Inspired Oscillatory Neural Networks for Edge Computing,” IEEE J. Emerg. Sel. Top. Circuits Syst., vol. 11, no. 4, pp. 586–596, Dec. 2021, doi: 10.1109/JETCAS.2021.3128756. [2] G. Indiveri and S.-C. Liu, “Memory and Information Processing in Neuromorphic Systems,” Proc. IEEE, vol. 103, no. 8, pp. 1379–1397, Aug. 2015, doi: 10.1109/JPROC.2015.2444094. [3] E. Corti, B. Gotsmann, K. Moselund, A. M. Ionescu, J. Robertson, and S. Karg, “Scaled resistively-coupled VO2 oscillators for neuromorphic computing,” Solid-State Electron., vol. 168, p. 107729, Jun. 2020, doi: 10.1016/j.sse.2019.107729. [4] E. Chicca, F. Stefanini, C. Bartolozzi, and G. Indiveri, “Neuromorphic Electronic Circuits for Building Autonomous Cognitive Systems,” Proc. IEEE, vol. 102, no. 9, pp. 1367–1388, Sep. 2014, doi: 10.1109/JPROC.2014.2313954. [5] G. Csaba and W. Porod, “Coupled oscillators for computing: A review and perspective,” Appl. Phys. Rev., vol. 7, no. 1, Mar. 2020, doi: 10.1063/1.5120412. [6] O. Maher et al., “A CMOS-compatible oscillation-based VO2 Ising machine solver,” Nat. Commun., vol. 15, p. 3334, 2024, doi: https://doi.org/10.1038/s41467-024-47642-5. [7] H. Mostafa, L. K. Müller, and G. Indiveri, “Rhythmic Inhibition Allows Neural Networks to Search for Maximally Consistent States,” Neural Comput., vol. 27, no. 12, pp. 2510–2547, Dec. 2015, doi: 10.1162/NECO_a_00785. [8] A. Raychowdhury et al., “Computing With Networks of Oscillatory Dynamical Systems,” Proc. IEEE, vol. 107, no. 1, pp. 73–89, Jan. 2019, doi: 10.1109/JPROC.2018.2878854. [9] O. Maher, N. Harnack, G. Indiveri, M. Sousa, B. Gotsmann, and S. Karg, “Solving optimization tasks power-efficiently exploiting VO2 ’s phase-change properties with Oscillating Neural Networks,” in 2023 Device Research Conference (DRC), IEEE, Jun. 2023, pp. 1–2. doi: 10.1109/DRC58590.2023.10186951. [10] O. Maher et al., “Highly Reproducible and CMOS-compatible VO2-based Oscillators for Brain-inspired Computing,” Sci. Rep., vol. 14, p. 11600, 2024, doi: https://doi.org/10.1038/s41598-024-61294-x. [11] E. Corti, C. Delacour, A. Todri-Sanial, and S. Karg, “Frequency Injection Locking-Controlled Oscillations for Synchronized Operations in VO2 Crossbar Devices,” in 2021 Device Research Conference (DRC), IEEE, Jun. 2021, pp. 1–2. doi: 10.1109/DRC52342.2021.9467129. [12] M. Desroches, B. Krauskopf, and H. M. Osinga, “Mixed-mode oscillations and slow manifolds in the self-coupled FitzHugh-Nagumo system,” Chaos Interdiscip. J. Nonlinear Sci., vol. 18, no. 1, Mar. 2008, doi: 10.1063/1.2799471.