Missions

Electronic accelerators currently powering artificial intelligence (AI) face intrinsic limitations, notably high-power consumption and latency, hindering performance in edge computing applications where efficiency and adaptability are critical [1]. With exponentially increasing data-driven tasks, energy-efficient AI computation becomes a strategic imperative, prompting exploration of alternative computing paradigms beyond conventional CMOS technology.
Photonic accelerators offer substantial advantages over traditional electronic circuits, leveraging the intrinsic parallelism and ultrafast dynamics of light to execute computations with high speed (tens of GHz) and low-power requirements [2]. Despite these benefits, training photonic neural networks remains challenging due to incompatibility with traditional digital training methods like backpropagation (BP), which necessitate energy-intensive digital simulation of gradients.
To address this critical limitation, this proposal investigates Equilibrium Propagation (EP), a biologically plausible training approach that naturally aligns with the dynamics of physical, analog systems. EP exploits system equilibrium states for gradient estimation without the explicit backward propagation steps of traditional BP, significantly reducing computational overhead and energy consumption.
This research focuses on three core objectives: (1) innovative photonic device concepts suitable for EP-based neural networks, (2) scalable photonic architectures specifically designed to accommodate EP dynamics leveraging the novel devices proposed, and (3) EP-based training efficacy and robustness on photonic hardware by means of comprehensive system-level simulations

Background and State-of-the-Art
Photonic systems have recently demonstrated their potential as effective AI accelerators, performing GHz-scale matrix operations with ultralow latency and near-digital precision (~8-bit accuracy) [3]. Universal photonic processors have been shown to implement complex AI models, such as convolutional neural networks (CNNs) and Transformers, with accuracy competitive to electronics [4]. However, existing approaches rely predominantly on offline training through digital BP, which imposes substantial energy overhead and mismatch issues when transitioning models from simulation to hardware.
EP offers a compelling alternative, using equilibrium states of physical systems to intrinsically compute gradients. Recently, EP-based algorithms implemented on analog electronic and quantum systems have demonstrated comparable accuracy to BP, confirming the feasibility of training neural networks entirely in hardware [5]. Complementing EP, forward-only photonic training schemes have shown promising results, demonstrating BP-equivalent accuracy without explicit feedback mechanisms [6]. These methods collectively suggest a viable pathway toward fully integrated, real-time learning within photonic neural systems.

Activités

Photonic Device Simulation
The proposal will explore, through extensive simulations, novel photonic device concepts optimized for EP-compatible learning. Devices will include tunable nonlinear optical components (such as microring resonators, semiconductor optical amplifiers, and electro-optic modulators) tailored to provide the dynamic nonlinearity and tunability required by EP-driven neural network operation. Simulations will focus on device-level optimization, aiming for high-speed reconfigurability, low power consumption, and sufficient precision for reliable in-hardware training. Various technological platforms will be considered such as SOI, SiN, and LNOI. Approaches where the photonic hardware is coupled with electronic hardware will be also considered to leverage the advantages of both technologies.

Scalable Photonic Architectures
Architectural development will leverage wavelength division multiplexing (WDM) and coherent interference-based computing paradigms. Simulations will validate scalable topologies capable of implementing recurrent neural network dynamics (e.g., through optical feedback integrated at a device or system-level [7]) essential to EP, allowing the system to autonomously reach equilibrium states and execute gradient-based weight updates fully optically. The architecture will integrate simulated coherent linear and nonlinear operations, ensuring realistic feasibility studies before hardware implementation.
Training Algorithms
This research will conduct comprehensive simulations of EP training on the proposed photonic architectures, demonstrating its suitability for autonomous gradient estimation within physical photonic systems. Simulations will evaluate performance metrics, including accuracy, convergence speed, and robustness to noise and device variability. Hybrid analog-digital training schemes will also be explored through simulations to enhance accuracy and adaptability while preserving computational efficiency. In particular, there'll be an analog training carried out at the photonic chip level, aided by a digital training to improve computing accuracy.

Compétences

Expected Outcomes and Impact
The research will deliver: (1) Advanced simulated designs of photonic devices optimized for EP learning, (2) Validated scalable photonic neural accelerator architectures, and (3) Demonstrated efficacy of EP-based training through rigorous simulations, showing performance comparable to BP but with significantly reduced energy requirements.
This integrated simulation-based research approach addresses fundamental barriers to scalable, efficient AI hardware, laying critical groundwork for future experimental realization. Such advances hold substantial promises for real-time adaptive AI applications, from edge computing and autonomous robotics to sustainable AI solutions. Prototypes through e.g., MPW runs could be foreseen depending on the degree of maturity of the results.

Contexte de travail

About the Project
This position is within the framework of the PEPR IA (project Emergences) : https://www.pepr-ia.fr/en/projet/emergence-2/ which aims at advance the state-of-the-art in emerging physics-based models by collaboratively exploring various computational models using the properties of different physical devices. The project focuses on bio-inspired event-driven models, physics-inspired models and innovative physics-based machine learning solutions. Emergences also intends to extend collaborative research activities beyond the consortium's perimeter, in conjunction with other PEPR projects and beyond other laboratories.

About the Institutes
The institutes involved in this position are the Center for Radiofrequencies, Optics, and Microelectronics in the Alps (CROMA), the Institute of Nanotechnologies of Lyon (INL), and the laboratory SPINTEC, part of the Center for Atomic and Renewable Energies – Institute for Interdisciplinary Research of Grenoble (CEA-IRIG).

About the Position
The physical location of the position is at the CROMA laboratory (Grenoble site) with regular interaction with the other institutes involved in the project. The duration of the position is for 2.5y.

References
[1] Y. Shen et al., "Deep learning with coherent nanophotonic circuits," Nat. Phot. 11, 441-446 (2017) : https://doi.org/10.1038/nphoton.2017.93
[2] B. Shastri et al., " Photonics for artificial intelligence and neuromorphic computing," Nat. Photon. 15, 102–114 (2021) : https://doi.org/10.1038/s41566-020-00754-y
[3] S. Hua et al., "An integrated large-scale photonic accelerator with ultralow latency," Nature 640, 361–367 (2025) : https://doi.org/10.1038/s41586-025-08786-6
[4] J. Feldmann et al., "Parallel convolution processing using an integrated photonic tensor core," Nature 589, 52–58 (2021) : https://doi.org/10.1038/s41586-020-03070-1
[5] J. Laydevant et al., "Training an Ising machine with equilibrium propagation," Nat. Comms 15, 3671 (2024) : https://doi.org/10.1038/s41467-024-46879-4
[6] S. Bandyopadhyay et al., "Single-chip photonic deep neural network with forward-only training," Nat. Photon. 18, 1335–1343 (2024) : https://doi.org/10.1038/s41566-024-01567-z
[7] M. Abdalla et al., "Minimum complexity integrated photonic architecture for delay-based reservoir computing, " Opt. Express 31, 7, 2023, https://doi.org/10.1364/OE.484052

Le poste se situe dans un secteur relevant de la protection du potentiel scientifique et technique (PPST), et nécessite donc, conformément à la réglementation, que votre arrivée soit autorisée par l'autorité compétente du MESR.