Neuromorphic Polariton Accelerator

Project Information

PolArt

Grant agreement ID: 101130304

DOI

10.3030/101130304

EC signature date 21 November 2023

Start date 1 February 2024

End date 31 January 2028

Funded under

• The European Innovation Council (EIC)

Total cost

€ 2 997 641,25

EU contribution

€ 2 997 641,25

Coordinated by

UNIWERSYTET WARSZAWSKI
Poland

Periodic Reporting for period 1 - PolArt (Neuromorphic Polariton Accelerator)

Reporting period: 2024-02-01 to 2025-01-31

Summary of the context and overall objectives of the project

The PolArt project focuses on the hardware implementation of a polariton-based neural network, which aims to accelerate computing tasks with unprecedented performance and reduced power consumption, as well as to ensure compatibility with standard electronics. We investigate its applications in three key areas: image recognition, sound recognition, and the detection of genome-wide biomarker patterns. For the first two tasks, this neuromorphic hardware accelerator is expected to significantly enhance processing speed compared to conventional computation relying solely on software-based convolutional neural network simulations. The third application is aimed at drastically reducing the time needed for genomic analysis using microarrays, which are playing an increasingly important role in clinical diagnostics, prognostics, and drug development.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

The first year of the PolArt project focused on designing and optimizing optical processors while addressing fabrication constraints and material nonlinearities. Collaboration between theoretical and experimental teams ensured that practical limitations were considered in developing polariton processor architectures. Two primary approaches were explored: binarized feed-forward and black-box architectures. Binarized neural networks were investigated as an energy-efficient alternative to analog networks, encoding activations and weights in binary form to reduce computational complexity. To support real-world implementation, advanced logic gates such as Toffoli and Feynman gates were considered, alongside an optimal pixel selection algorithm to enhance scalability. A new polariton-based Toffoli-like gate was introduced, using a three-arm polariton coupler with phase-encoded input information. The black-box neural network approach employed surrogate models to approximate complex physical interactions where direct measurements are difficult.
On the experimental side we developed a scalable method for fabricating CsPbBr3 microwire waveguides that support room-temperature polariton condensation and lasing. Using a microfluidic-assisted process with PDMS templates, high-quality single-crystal structures were formed. Optical characterization confirmed strong exciton-photon coupling and long-range polariton propagation. The method's simplicity and scalability make it promising for on-chip photonics. We further extended the application of our perovskite waveguides and proposed a room-temperature optical neural network. The exciton-polariton condensation in waveguides was used as the nonlinear activation function resembling the widely used ReLU in machine learning. Four polariton neurons were demonstrated to successfully classify objects with 96% accuracy and solved complex binary classification problems with over 92% accuracy, significantly outperforming linear classifiers. The low energy consumption of 175 pJ per operation highlights the potential of this platform for scalable, high-efficiency optical computing.
Our successful fabrication of GaAs-based, GaN-based and perovskite waveguides, and the design of integration with GaN microlasers and micro-LED arrays demonstrate significant potential for neuromorphic optical processing. The design of integrated waveguide matrices and nonlinear couplers marks a crucial step toward scalable photonic neural networks. Future work will focus on optimizing perovskite growth, refining DBR structures, and improving optical components to advance the realization of an integrated photonic computing platform.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

To ensure the further uptake and success of these technologies, additional research and demonstration efforts are necessary. Key priorities include scaling polariton-based neural networks to multi-layer and large-scale architectures, optimizing material growth for improved stability and propagation distances, and refining waveguide structures to enhance efficiency. Access to markets and finance is critical for commercialization. Collaboration with semiconductor and AI industries will support the development of commercialization strategies for polariton-based neuromorphic chips. Securing funding will facilitate the transition from research prototypes to scalable AI and biomedical applications.
Intellectual property protection is essential, particularly for polariton neural networks, integrated photonic circuits, and micro-LED-perovskite/GaN microlasers. Identifying patentable innovations and developing technology transfer strategies will enable engagement with chip manufacturers, AI hardware developers, and biophotonics companies.
Strengthening partnerships with leading research institutions, semiconductor industries, and AI hardware companies will accelerate technological progress. Additionally, establishing links with European and international photonics and quantum technology initiatives will create opportunities for funding, industrial collaboration, and regulatory alignment to support standardization and market adoption.