Periodic Reporting for period 1 - NEUROPIC (Nano electro-optomechanical programmable integrated circuits)
Reporting period: 2023-03-01 to 2024-02-29
Artificial intelligence relies on training algorithms to process large amounts of data in parallel from which we compute new information. Training and computing these types of algorithms in standard sequential hardware consumes vast amounts of energy. To minimize this energy bill in a context where the ever-increasing use of this technology is a top-level scientific and technological priority. Different proposals to implement parallel-processing hardware exists using electronic systems which are the natural candidates to extent microelectronics to parallel computing. However, photonics offers a natural paradigm fully compatible with parallel computing which advantages are (1) massively parallel data transfer and (2) extremely high data modulation speeds limited only by the bandwidth of on-chip optical modulators and photodetectors.
NEUROPIC exploits silicon extreme nanomachining with massively-parallel integration and interconnection of photonic meshes where nanoelectromechanical components and optomechanical interaction are combined for reservoir computing at high speed. Our major goal is to implement an integrated reservoir where the single neuron is given the nonlinear nature of the optomechanical coupling at the nanoscale. NEUROPIC advanced technology has additional transformational impact potential on photonics for data centers, autonomous vehicles, quantum information processors, and much more.
NEUROPIC exploits silicon extreme nanomachining with massively-parallel integration and interconnection of photonic meshes where nanoelectromechanical components and optomechanical interaction are combined for reservoir computing at high speed. Our major goal is to implement an integrated reservoir where the single neuron is given the nonlinear nature of the optomechanical coupling at the nanoscale. NEUROPIC advanced technology has additional transformational impact potential on photonics for data centers, autonomous vehicles, quantum information processors, and much more.
During the first year, NEUROPIC has progressed in various of its major research lines. We have so far worked on characterizing and optimizing the electron-beam silicon lithography process which is at the basis of our technology. Our advances will have a clear impact not only on our project but on a technique used by a broad spectrum of users in research and industry. Extreme silicon fabrication pushing the spatial resolution of the footprint has been achieved only during this first year of the project [Self-assembled photonic cavities with atomic-scale confinement. AN Babar, TAS Weis, K Tsoukalas, S Kadkhodazadeh, G Arregui, et al. Nature 624 (7990), 57-63 (2023). An important part of the project relies on the modelling and design of electrostatic NEMS [https://arxiv.org/abs/2307.01122] in which a systematic direct comparison between experiment and theory for the displacement, operating bandwidth, and several other important parameters. These developments have enabled us to implement amplitude and phase modulators at low frequency so far (about few MHz). The consortium has already progressed on the specifications required for a neural network considering only switching network nodes and full-programable nodes. We have quantified the number of electrical and optical connections for each option alongside the fabrication capabilities within the consortium. Finally, we have optimized our designs of the single optomechanical neuron based on our own expertise and research collaboration between different partners of the consortium. We have developed an optimization analysis of the performance of the neuron, an optomechanical cavity formed by an optical resonance with a frequency around 200 THz and a mechanical resonance with a frequency around 10 GHz simultaneously confined within the same volume. The performance of the neuron has been maximized by maximizing both the quality factor of the optical resonance (Q) and the optomechanical coupling between the optical and the mechanical resonances (gOM).
During the first year of our project, we have already achieved significant progress beyond the state of the art mostly in our activities in silicon nanomachining. Part of these activities need to be so far protected and cannot be disclosed in a public summary, but some of them have already been published as in the Self-assembled photonic cavities with atomic-scale confinement. Ali N. B. et al. Nature volume 624, 57 (2023) where the DTU team shows impressive results pushing silicon electron-beam lithography to an extreme.