Combined spatially and temporally multiplexed photonic reservoir computer with a diffractively coupled VCSEL-array Preprint: AIP/123-QED

J. Robertson, M. Pflüger, I. Fischer, M. C. Soriano, A. Hurtado · 2026

Read Combined spatially and temporally multiplexed photonic reservoir computer with a diffractively coupled VCSEL-array Preprint: AIP/123-QED by J. Robertson, M. Pflüger, I. Fischer, M. C. Soriano, A. Hurtado (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-2605-00753.txt.

Introduction

Joshua Robertson and 4 co-authors report an experimental photonic reservoir computer that classifies data using an array of semiconductor lasers coupled through free space. Reservoir computing lets a fixed, complex physical system do the hard transformation, training only a simple readout. Their hybrid design multiplexes in space and time at once. The laser array supplies spatial nodes through diffractive coupling in an external cavity. Time multiplexing creates up to eighty-eight virtual nodes per laser.

Combining the two expansions beat either alone, driving the test error down to a few percent. A network of twelve spatial nodes was stretched into an effective network of nearly a thousand nodes. The input step took under eighteen nanoseconds, thanks to the laser transformation's bandwidth. Performance depended on both node counts, which the authors map for future builders.

Catalog record

CREATORJ. Robertson, M. Pflüger, I. Fischer, M. C. Soriano, A. Hurtado
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.00753

Browse the full library · About the Tower of Babel

Full text

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

Combined spatially and temporally multiplexed photonic reservoir computer with a diffractively coupled VCSEL-array Preprint: AIP/123-QED J. Robertson, M. Pflüger, I. Fischer, M. C. Soriano, A. Hurtado Abstract We report and analyse the classification performance of an experimental hybrid spatio-temporal photonic reservoir computer based upon a free-space VCSEL array. We demonstrate experimentally the enhancement of spatial-only reservoir operation, featuring the diffractive coupling of lasers in an external cavity, by exploiting up to 88 virtual nodes with time multiplexing. We analyse the dependance of performance on the spatial and virtual node number, and achieve an improvement for both spatial- and temporal-only reservoirs with a reduced test error of 0.026 in a classification task. Further, given the high bandwidth of the non-linear laser transformation, we demonstrate the expansion of a 12 spatial node network to a 968 node network, operating at an input time of 17.6 ns, maintaining high processing speed and improving network scalability and performance. I INTRODUCTION Reservoir computing (RC), an unconventional-computing paradigm for neural network implementation, is a promising route to alleviating the challenging energy efficiency and speed bottlenecks faced by Von Neumann computing architectures in the booming artificial intelligence (AI) era Yan et al. 2024 ; Nakajima and Fischer 2021 . Reservoir computers are used to create simplified but powerful recurrent neural networks that exploit a rich feature space, expanding information to a high dimensional representation using various non-linear transformations Maass, Natschläger, and Markram 2002 ; Jaeger and Haas 2004 . In RC, the network manifests as a ‘reservoir’ of nodes with fixed unknown connections and weights. This feature simplifies the learning (training) requirements to a small set of output node weights (the readout layer), improving the efficiency and speed of system training Lukoševičius and Jaeger 2009 . Catalysed by innovations making learning concepts hardware-agnostic, neuromorphic computing has been thriving. A core appeal of RC is therefore its deployability to physical hardware Tanaka et al. 2019 , and given the demand for energy efficient, high-speed computing, the photonic platform is being actively researched Van der Sande, Brunner, and Soriano 2017 ; Abreu et al. 2024 . This is giving rise to the field of Photonic Reservoir Computing (PRC) Abdalla et al. 2025 . Large bandwidth, low-loss passive components (e.g. waveguides and couplers), high-speed low-latency transmission and access to optical nonlinear elements is allowing for the photonic platform to create promising all-optical RC systems void of costly optical-electrical conversions. The first demonstrations of PRC typically deployed semiconductor lasers Brunner et al. 2013 , and optical amplifiers Duport et al. 2012 , but now approaches based on photonic integrated circuit (PIC) components, such as microring resonators, have been demonstrated, highlighting the flexibility, versatility and scalability of the platform and PRC techniques Vandoorne et al. 2014 ; Donati et al. 2024 ; Donati et al. 2025 . PRC architectures typically fall into one of two species, spatial (physically networked) or temporal (delayed-based) Abreu et al. 2024 . Temporal reservoirs leverage a single physical nonlinear node to create a network of many coupled virtual nodes via time-multiplexing Appeltant et al. 2011 . Coupling and memory in these systems is created through the short-term dynamics in the system, and the long-term feedback created through an optical delay line. Spatial systems, on the other hand, implement multiple discrete nonlinear nodes, that are physically coupled using various techniques (waveguides, diffractive optics, surface-interactions etc.), to create truly parallel reservoir nodes in 2D structures/arrays Rafayelyan et al. 2020 ; Antonik, Marsal, and Rontani 2019 ; Bueno et al. 2018 ; Pflüger et al. 2024 . Each architecture provides unique benefits; temporal reservoirs are more hardware-friendly, easily scalable, and serve well memory-intensive tasks; while spatial systems provide instantaneous input transformation for real-time operations. The combination of spatial and temporal architectures has also been proposed in recent theoretical studies Röhm and Lüdge 2018 ; Bauwens et al. 2024 ; Huang et al. 2023 . These systems aim to compliment the high virtual node density of temporal reservoirs with the distributed complexity and inherent parallelism (speed) of spatial reservoirs. Recent studies include modeling of a spatial PIC reservoir coupled to a delayed-based semiconductor laser reservoir Bauwens et al. 2024 ; and a study on multiple temporal VCSEL reservoirs in parallel, with mutual coupling configurations Huang et al. 2023 . On the experimental side, reservoirs employing both wavelength and temporal multiplexing for high throughput and computational efficiency has also been demonstrated Lupo et al. 2023 ; Aadhi et al. 2025 . However, beyond the time-multiplexing of matrix vector multiplication operations for deep neural networks Chen et al. 2023 , reports of experimental spatio-temporal reservoir architectures with an array of lasers is lacking. In this work, an experimental reservoir with both spatial and temporal multiplexing (spatio-temporal) is proposed based upon an array of diffractively-coupled VCSELs. The proposed spatio-temporal PRC system uses Iris flower classification as a benchmarking task, before the performance is analysed for increasing spatial and virtual (temporal) node counts. This experimental work highlights that the performance of an already existing PRC can be enhanced using a hybrid operation approach, via the introduction of additional virtual nodes. This approach benefits from the fast dynamics of semiconductors lasers, imposing only a small trade-off in operation speed. The following Sections II & III , introduce the experimental array-based PRC and present its classification performance for different configurations, respectively. II EXPERIMENTAL SETUP: SPATIO-TEMPORAL VCSEL ARRAY PRC In this work, the core non-linear photonic elements are VCSELs of a custom-manufactured VCSEL array, previously introduced in Heuser et al. 2020 ; Pflüger et al. 2024 ; Pflüger et al. 2023 . The VCSELs of the array are arranged in a 5\times 5 square lattice with a pitch of p\approx 80\,{\mu m} between neighbouring VCSELs. Each laser is electrically contacted allowing their pump current to be controlled independently. The VCSELs incorporate AlGaAs/GaAs Distributed Bragg Reflector mirrors and an active region of GaInAs quantum wells with GaAsP barrier layers. The aperture of the lasers are defined by an oxidised AlGaAs layer in the previously described mesa. The lasers emit with a dominant fundamental transverse mode at \sim 976 nm with a high degree of homogeneity in their polarisation orientation due to a slight elliptical cross-section. More fabrication details can be found in Heuser et al. 2020 . For individual VCSELs, we define a nomenclature as shown in Fig. 1 . A VCSEL in column X and row Y of the array is referred to as VCSEL (cX,rY) . Figure 2 shows the experimental setup of the array-based PRC and the proposed implementation of both spatial and virtual nodes. This setup uses light from an external injection laser (inj, Thorlabs DBR976PN), that passes a Mach-Zehnder modulator (MZM, EOSpace AZ-0K5-10-PFA-PFA-970), to input intensity-modulated information into the system. The injection light is emitted from a polarisation-maintaining fiber tip and is collimated using an aspheric lens (L1), before it is polarisation-aligned to the VCSEL emission using a \lambda/2 waveplate. An external feedback cavity, created using a microscope objective, a lens (L3) and a mirror is used to introduce recurrence and spatial coupling between VCSELs via a diffractive optical element (DOE). The injection enters the external cavity via a reflection at BS1 and passes the DOE twice before being injected into the VCSELs. This

Read the complete text