Reconfigurable Diffractive Optical Neural Networks with Phase Change Material based Photonic Device

衍射光神经网络（DONNs）系统作为执行机器学习任务的高性能光学架构已经引起了人们的兴趣。对于理想的DONNs系统，目前还缺乏节能的衍射像元单元和精确的软件模型。该项目采用一种称为相变材料（PCM）的非易失性材料，并解决了两大挑战，即大量的开关能量和多电平操作，以开发基于相变材料的衍射器件。本项目还通过结合层间和层内效应建立了一个精确的模型。本课题的研究成果在通信、计算、量子等领域具有广泛的光子与光电应用前景。该项目还通过在实验室、课堂和外展项目开展培训和教育活动，扩大学生对科学、技术、工程和数学（STEM）的参与。这些活动的目标是培养多样化的未来STEM劳动力。DONNs系统通过空间光调制和多个衍射层的光学衍射来执行机器学习任务。然而，为了实现DONNs系统的最终全光、完全可重构和紧凑的衍射层，存在技术差距，包括非易失性可重构性，以及准确和可训练的软件模型。为了填补这些空白，该项目采用具有一些理想特性的非易失性硫系pcm，如内存计算、大光学对比度和具有高循环性的超快重构，为DONNs系统构建近红外衍射器件。本项目旨在解决以下挑战，包括实现基于pcm的光子器件的大重构能耗和多能级运行，以及标准DONNs模型与基于pcm的衍射器件的紧凑DONNs系统之间的差异。具体来说，该项目创造了一种节能透明的电加热器，用于重新配置pcm，使用排列整齐的碳纳米管薄膜，具有非凡的、可单独优化的电学、热学和光学特性。本项目还设计、优化和制造了一个多层可重构器件，该器件仅在多个PCM薄膜中使用两种可靠的晶体和非晶态。此外，本项目通过结合层间反射和层内像素间相互作用的影响，实现了精确和可训练的donn模型。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Diffractive optical neural networks (DONNs) systems have gained interest as high-performance optical architectures to perform machine learning tasks. Toward the ideal DONNs systems, there is a lack of energy-efficient diffractive pixel unit and accurate software models. This project employs one type of nonvolatile material called phase change material (PCM) and address two major challenges, substantial switching energy and multilevel operations, to develop PCM-based diffractive devices. This project also develops an accurate model by incorporating interlayer and intralayer effects. The research findings from this project can find broad photonic and optoelectronic applications, such as in communication, computation, and quantum technologies. This project also expands participation in science, technology, engineering, and math (STEM) through training and education activities in the laboratory, classroom, and through outreach programs. The goal of these activities is to develop a diverse future STEM workforce.DONNs systems perform machine learning tasks through spatial light modulation and optical diffraction in multiple diffractive layers. However, toward the implementation of the ultimate all-optical, fully reconfigurable, and compact diffractive layers for DONNs systems, there exist technological gaps including nonvolatile reconfigurability, and accurate and trainable software models. To fill these gaps, this project employs nonvolatile chalcogenide PCMs that feature a few desirable properties, such as in-memory computing, large optical contrast, and ultrafast reconfiguration with high cyclability, to construct a near-infrared diffractive device for DONNs systems. This project aims to address following challenges, including large reconfiguration energy consumption and multilevel operation for implementing PCM-based photonic devices, as well as the discrepancy between the standard DONNs model and the compact DONNs system with PCM-based diffractive devices. Specifically, this project creates an energy-efficient and transparent electrical heater for reconfiguring PCMs using aligned carbon nanotube films with extraordinary and separately optimizable electrical, thermal, and optical properties. This project also designs, optimizes, and fabricates a multilevel reconfigurable device by only using two reliable crystalline and amorphous states in multiple PCM films. In addition, this project implements accurate and trainable DONNs models by incorporating the effects of interlayer reflection and intralayer interpixel interaction.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.