Neuromorphic computing with optomechanical oscillators

Andrea Gaspari, Rémi Avriller, Florian Marquardt, Fabio Pistolesi · 2026

Read Neuromorphic computing with optomechanical oscillators by Andrea Gaspari, Rémi Avriller, Florian Marquardt, Fabio Pistolesi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-11658.txt.

Introduction

Gaspari and three colleagues study brain-like computing with optomechanical oscillators. Phase oscillators suit machine learning through built-in nonlinearity and joint sync when coupled. The authors study a network of oscillators pumped in the blue-detuned regime for self-sustained oscillations, propose a theory for their motion, and show how such systems serve brain-like computing. They discuss training and a drum-resonator platform for a real build. Results from modeling an XOR gate with 5 nodes in an all-to-all setup round out the work.

Catalog record

CREATORAndrea Gaspari, Rémi Avriller, Florian Marquardt, Fabio Pistolesi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE80066
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.11658

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Neuromorphic computing with optomechanical oscillators Andrea Gaspari, Rémi Avriller, Florian Marquardt, Fabio Pistolesi Abstract The increasing resource demands of artificial neural networks have prompted the exploration of novel platforms better suited for machine learning. In this context, phase oscillators represent a promising candidate due to their intrinsic nonlinearity and their ability to exhibit collective synchronization when coupled together. In the present work, we investigate one such implementation: a network of optomechanical oscillators pumped in the blue–detuned regime to achieve self–sustained oscillations. We propose a theoretical framework to describe their dynamics and demonstrate how such systems can be employed for neuromorphic computing. We discuss how they can be trained and analyze a platform, based on drum resonators, that could enable their physical implementation. Ultimately, the theoretical results obtained from modelling an XOR gate using 5 nodes in an all–to–all configuration are discussed. I Introduction Artificial neural networks 21 (ANNs) have seen remarkable progress in recent years, driven by advances in both theoretical foundations and practical implementations. These developments have enabled models of unprecedented scale and accuracy, capable of tackling highly complex tasks. At the same time, this rapid growth has introduced significant engineering challenges arising from the increasing number of nodes and trainable parameters. For state–of–the–art ANNs, solving the network dynamics and implementing training protocols on conventional von Neumann computers 39 have become extraordinarily energy–intensive, exposing fundamental limitations. Computers, as deterministic and error-free platforms, are fundamentally inefficient at executing inherently stochastic machine learning algorithms, leading to a waste of computing power. Additionally, the physical separation between processing and memory units necessitates constant data transfer, which consumes substantial resources, especially in large–scale networks. This architecture also introduces a bottleneck effect: when computational demands exceed the data transfer capacity, the entire process slows down. In this context, attention has increasingly focused on finding new solutions able to reduce energy consumption, mitigate the von Neumann bottleneck and enhance parallelism. The field of neuromorphic computing 5 ; 25 seeks to address these challenges by proposing novel platforms better suited to machine learning. Notably, one of the main strategies adopts a genuinely physical approach, aiming to replace artificial neurons with dynamical systems and harness their intrinsic nonlinear behavior for computation. Many implementations have been already explored, including networks based on memristors 30 , spin oscillators 38 ; 31 and Josephson junctions 33 , as well as optical proposals involving light scattering in complex media 10 ; 40 , multiple light scattering 44 , multiple–scattering cavities 43 , linear wave scattering 42 , diffractive optical elements 3 and coherent nanophotonic circuits 34 . However, it remains unclear which physical platforms are most suitable in terms of efficiency, scalability and expressivity. This motivates the exploration of further system classes— ideally those with established communities and robust theoretical foundations. Following the remarkable results obtained using phase oscillators 41 , our attention was directed toward optomechanical implementations, broadly defined across the microwave to the optical regime. These systems exhibit a broad range of properties 1 that make them particularly attractive for neuromorphic computing. Specifically, they feature an intrinsic nonlinearity arising from the interaction with the cavity field (essential for nonlinear computing), micro– and nanoscale physical dimensions, high quality factors and high oscillation frequencies, as well as the possibility of being cooled down to millikelvin temperatures to suppress thermal noise. Moreover, they offer precise tunability of the fabrication parameters, making them extremely versatile platforms for experimental implementations. In parallel, when the mechanical damping is balanced by a pumping source, these systems enter the self-sustained regime and can be effectively treated as phase oscillators. Under such condition, they undergo a Hopf bifurcation and start oscillating on stable limit orbits. Networks made of phase oscillators are particularly appealing for computing as they can display collective phenomena like synchronization 29 — an emergent behavior, well documented in neuroscience, which we aim to exploit. The structure of the paper is as follows. Starting from the theoretical model, in Sec. II.1 and II.2 the individual and collective dynamics of optomechanical oscillators is analyzed, developing a framework tailored for neuromorphic computing. The approach relies on using oscillator phases as node variables and leverages coupling, external driving and synchronization to reproduce the main features of artificial neural networks. In Sec. III.1 , the implementation of inference and training on physical platforms is presented, with a particular focus on the role of dynamical multistability discussed in Sec. III.2 . To complete the analysis, in Sec. III.3 we present a model to physically realize the interconnections using drum resonators 36 ; 28 while in Sec. III.4 we delve into the case of SiN drum resonators 47 , providing parameter ranges for the experimental realization. Ultimately, the entire scheme is validated through the theoretical implementation of an XOR gate (Sec. IV ), an ideal benchmark in machine learning due to its simplicity. The results, collected in Sec. IV.3 , confirm the network’s expressivity, i.e. its ability to successfully solve the task. II Model Starting from the physical description of a single optomechanical oscillator, this section gradually builds the full network model, placing particular emphasis on the analogies with ANNs. II.1 Optomechanical nodes Optomechanical oscillators are systems in which a cavity mode is coupled to a mechanical degree of freedom. The canonical setup is that of an optical Fabry–Pérot cavity with a movable end mirror, Fig. 1 -a, yet the general framework is quite broad and also encompasses implementation based on electrical circuits, Fig. 1 -b. In Hamiltonian formalism, these systems are described as two coupled quantum harmonic oscillators. Using \hat{a} and \hat{b} to denote the annihilation operators for the cavity and the mechanical mode, respectively, their interaction emerges from the Taylor expansion of the cavity resonance frequency \omega_{c} with respect to the mechanical displacement \hat{x}=x_{\mathrm{zpf}}(\hat{b}^{\dagger}+\hat{b}) , with x_{\mathrm{zpf}}=\sqrt{\hbar/2m\Omega} , the zero point fluctuations. This yields the optomechanical Hamiltonian 1 : \hat{H}=\hbar\omega_{c}\,\hat{a}^{\dagger}\hat{a}+\hbar\Omega\,\hat{b}^{\dagger}\hat{b}-\hbar G(\hat{a}^{\dagger}\hat{a})\hat{x}. (1) The emerging radiation pressure force, \hat{F}=-\partial\hat{H}/\partial\hat{x}=\hbar G(\hat{a}^{\dagger}\hat{a}) , acts on the mechanical mode resulting in a shift of the optical frequency \omega_{c} , directly proportional to the mechanical displacement through G , the frequency pull parameter. For the classical cavity field amplitude, \alpha=\langle\hat{a}\rangle , and mechanical oscillator displacement, x=\langle\hat{x}\rangle , the dynamics— including dissipation sources— is described by the following set of coupled equations 1 : \begin{split}&\dot{\alpha}=\Big[i(\Delta+Gx)-\frac{\kappa}{2}\Big]\alpha+\frac{\kappa}{2}\alpha_{\mathrm{max}},\\ &m\ddot{x}=-m\Omega^{2}x-m\Gamma\dot{x}+\hbar G|\alpha|^{2}.\end{split} (2) Here, \Delta=\omega_{l}-\omega_{c} , \kappa , and \alpha_{\mathrm{max}} are the detuning of the laser source from the natural resonance frequency, the cavity decay rate, and the maximum light–field amplitude, respectively. The parameters m,\Omega and \Gamma correspond to the mass, the frequency and the intrinsic damping of the mechanical oscillator. Figure 1: Common examples of cavity optomechanics setups, highlighting the role of the cavity and the mechanical mode. (a) Cavity optomechanics setup consisting of a Fabry–Pérot cavity with a movable end mirror. (b) Cavity optomechanics analogue using electrical components. The physics resulting from this coupled system is remarkably rich: when the laser is red–detuned, \Delta<0 , the mechanical mode is cooled down, whereas blue detuning, \Delta>0 , leads to negative damping 16 . These processes are captured by the optomechanical induced damping \Gamma_{\mathrm{Opt}} , which takes positive or negative values, respectively. When blue–detuned, if: \Gamma+\Gamma_{\mathrm{Opt}}<0, (3) the system enters the nonlinear regime and begins to exhibit self-sustained oscillations, effectively turning

Read the complete text