Room temperature exciton-polariton neural network with perovskite crystal

Andrzej Opala 1,2,†, Krzysztof Tyszka 1,†, Mateusz Kędziora 1, Magdalena Furman 1, Amir Rahmani 2, Stanisław Świerczewski 1, Marek Ekielski 3, Anna Szerling 3, Michał Matuszewski 2,4 and Barbara Piętka 1,∗ · 2024

Read Room temperature exciton-polariton neural network with perovskite crystal by Andrzej Opala 1,2,†, Krzysztof Tyszka 1,†, Mateusz Kędziora 1, Magdalena Furman 1, Amir Rahmani 2, Stanisław Świerczewski 1, Marek Ekielski 3, Anna Szerling 3, Michał Matuszewski 2,4 and Barbara Piętka 1,∗ (2024) 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-2412-10865.txt.

Introduction

Andrzej Opala, Krzysztof Tyszka, and eight co-authors demonstrate, in a 2024 preprint, a neural network made of light and matter that works at room temperature. The network's units are exciton-polaritons, hybrid particles that form when light couples strongly to matter. They interact strongly and move fast, which suits computing. Every earlier polariton network, though, needed deep cooling to run. That kept the idea locked in the lab.

The team used a single-crystal perovskite waveguide to hold a polariton condensate without cooling. The device then carried out machine learning tasks. It sorted binary inputs correctly and detected objects in test cases. The authors call the result a milestone for optical computing. Room-temperature operation opens a path toward fast, low-energy accelerators. Perovskite materials, cheap and easy to shape, may carry that path forward.

Catalog record

CREATORAndrzej Opala 1,2,†, Krzysztof Tyszka 1,†, Mateusz Kędziora 1, Magdalena Furman 1, Amir Rahmani 2, Stanisław Świerczewski 1, Marek Ekielski 3, Anna Szerling 3, Michał Matuszewski 2,4 and Barbara Piętka 1,∗
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-10865

Browse the full library · About the Tower of Babel

Full text

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

Room temperature exciton-polariton neural network with perovskite crystal Andrzej Opala 1,2,†, Krzysztof Tyszka 1,†, Mateusz Kędziora 1, Magdalena Furman 1, Amir Rahmani 2, Stanisław Świerczewski 1, Marek Ekielski 3, Anna Szerling 3, Michał Matuszewski 2,4 and Barbara Piętka 1,∗ Abstract Limitations of electronics have stimulated the search for novel unconventional computing platforms that enable energy-efficient and ultra-fast information processing. Among various systems, exciton-polaritons stand out as promising candidates for the realization of optical neuromorphic devices. This is due to their unique hybrid light-matter properties, resulting in strong optical nonlinearity and excellent transport capabilities. However, previous implementations of polariton neural networks have been restricted to cryogenic temperatures, limiting their practical applications. In this work, using non-equillibrium Bose-Einstein condensation in a monocrystalline perovskite waveguide, we demonstrate the first room-temperature exciton-polariton neural network. Its performance is verified in various machine learning tasks, including binary classification, and object detection. Our result is a crucial milestone in the development of practical applications of polariton neural networks and provides new perspectives for optical computing accelerators based on perovskites. {spacing} 1.125 {affiliations} Institute of Experimental Physics, Faculty of Physics, University of Warsaw, ul. Pasteura 5, PL-02-093 Warsaw, Poland Institute of Physics, Polish Academy of Sciences, Aleja Lotników 32/46, PL-02-668 Warsaw, Poland Łukasiewicz Research Network - Institute of Microelectronics and Photonics, al. Lotników 32/46, 02-668, Warsaw, Poland Center for Theoretical Physics, Polish Academy of Sciences Aleja Lotników 32/46, 02-668 Warsaw, Poland {\dagger} - equally contributed {*} - corresponding author: [email protected] Neuromorphic computing systems (NCSs) are hardware-based computing devices inspired by the structure and function of biological neurons. Unlike traditional software-based ANNs, where neurons are simulated by algorithms [ 1 ] , neurons in NCSs are implemented directly in hardware. This allows for the simultaneous storage and processing of information, resulting in computational parallelism and overcoming the limitations of Turing-type electronics. In particular, photonic NCSs use optical elements to process information encoded in the degrees of freedom of light, including intensity, polarization, and phase [ 2 ] . Over the last few decades, realizations of neuromorphic devices based on different optical platforms, such as nanophotonic circuits [ 3 ] , phase-change materials [ 4 ] , and disordered media [ 5 ] , among others [ 6 ] , have made significant progress [ 7 ] . Despite these advances, optical neuromorphic computing still faces challenges. A significant obstacle in creating nonlinear hardware neurons is the lack of photon-photon interaction. To address this problem, many optical hardware neural networks have adopted a hybrid optoelectronic approach, where optical elements handle information transfer while electronic components perform nonlinear processing. This hybrid solution combines the speed of optical processing with the flexibility of electronic systems, including tunable nonlinearity. However, this is a remedy only for one of the issues, since optoelectronic data conversion is itself a bottleneck limiting the ability to provide energy-efficient computing [ 8 ] . A promising alternative for the hybrid approach involves the use of exciton-polaritons (hereafter polaritons), which are half-light, half-matter bosonic quasiparticles providing stronger light-matter interactions [ 9 , 10 ] . Polaritons combine electronic and photonic properties in a single quasiparticle, forming a unified, optically controlled physical platform [ 11 , 12 ] . They are formed when the coherent exchange of energy between an elementary electronic excitation of a semiconductor crystal (exciton) and a quantum of the electromagnetic field (photon) occurs on a time scale significantly shorter than its lifetime, which is observed in the so-called strong coupling regime [ 13 , 14 ] . In these conditions, the two constituents lose their distinct nature and form a bosonic quasiparticle remaining in a coherent superposition. Polaritons exhibit bosonic statistics and are capable of undergoing a phase transition to a non-equilibrium bosonic condensate [ 15 ] even at room temperature [ 16 ] . The photonic component of polariton results in an ultra-low effective mass, leading to excellent transport properties, when the excitonic component of polaritons gives rise to significant interparticle interactions. From the perspective of applications, polaritons have been used for creating optical lattices [ 17 ] , simulators [ 18 , 19 ] , energy-efficient, ultrafast optical switches [ 20 , 21 ] , logic gates [ 22 , 23 , 24 ] , routers [ 25 ] , transistors [ 26 , 27 ] , and neural networks [ 9 , 10 ] . Nevertheless, neural network implementations have been limited to cryogenic temperatures. Therefore, achieving a polariton neural network that operates at room temperature remained one of the major challenges in the field of optical computing with exciton-polaritons. This work overcomes this challenge by using lead halide perovskites, an emerging semiconductor material, to realize the first polariton neural network operating at room temperature. In our implementation, we use separated polariton neurons whose activation is correlated with the transition to the non-equilibrium Bose-Einstein condensate. We demonstrate the performance of the proposed neuromorphic system in machine learning tasks such as object detection and binary classification. In the first task, we employ polariton NCS to distinguish images of different geometrical objects. The analyzed objects fell into four categories: small circles, large circles, squares, and empty planes. This task is particularly challenging for any linear classifier due to the unspecified object positions. However, based on experimental results and considering the real noise, we show that a polariton neural network with only four neurons, whose activation functions resemble a parametric rectified linear unit (ReLU), is able to solve this problem. We classified 4,000 objects with an accuracy of \sim 96 \% (experimental inference), higher than the average accuracy obtained with an ideal linear classifier, more than \sim 25 \% and lower than software prediction by only \sim 3 \% . These results underscore the impact of polariton high nonlinearity, induced by the possibility of condensation at room temperature, and the relevance of polariton neuromorphic systems in machine learning applications. Next, we validated the perovskite-based polariton neural network on a binary classification problem with linearly inseparable datasets. We achieved classification accuracies that outperformed the benchmark linear classifier accuracy by more than 92 \% for three different types of datasets with varying levels of complexity. 0.1 Room temperature exciton-polariton neural network. Polariton neurons are characterized by a strong nonlinearity, resulting from polariton condensation and strong interparticle interactions mediated by their excitonic component. Previous experimental implementations of polariton neural networks, inspired by the theoretical concepts [ 28 , 29 , 30 ] , have used optical signals modulated in space, time, or both, to activate polariton neurons. These systems have been successfully developed in both recurrent and feed-forward neural network architectures and have been tested on tasks such as handwritten digit and spoken word classification, demonstrating high energy efficiency [ 31 , 22 , 32 , 33 ] . The mentioned realizations of polariton NCS have been limited to optical microcavities based on gallium arsenide and cadmium telluride semiconductors, which exhibit strong nonlinear effects. However, because of the low exciton binding energies and low exciton-photon coupling strength, the applicability of these materials is typically limited to cryogenic temperatures. Finding a solution to this problem by harnessing emerging semiconductor materials was pointed out as a crucial step to achieving real-world applications of polariton neuromorphic system [ 34 , 10 ]

Read the complete text