Scholarlink

<p>Inverse design of on-chip nanophotonic neural networks (ONNs) is often hindered by the prohibitive cost of repeated three-dimensional (3D) full-wave simulations, especially when the optimization is performed in a high-dimensional geometric design space. Here, we employ a computationally efficient multi-body shape optimization framework and integrate it with a wave-based linear-decoupling strategy to address this challenge. By reconstructing input-specific responses from pre-computed characteristic states, our approach bypasses the need for exhaustive, repetitive full-wave solutions of Maxwell’s equations inside the optimization loop. We utilize elliptical scatterers to enable anisotropic control over optical fields, reducing the nominal design-variable count by hundreds of times for the studied devices compared to an equivalent 40-nm pixel-based topology-optimization baseline. 3D finite-difference time-domain simulations on the Iris and MNIST classification benchmarks validate that the proposed framework yields compact ONN classifiers with high inference accuracy and significantly lower training cost. This work establishes a scalable, fabrication-aware methodology for high-performance on-chip photonic neural processors.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient multi-body shape optimization of on-chip nanophotonic neural network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月16日

查看全文
  • Licheng Wang,
  • Qiwen Bao,
  • Xianjin Liu,
  • Jun-Jun Xiao

摘要

Inverse design of on-chip nanophotonic neural networks (ONNs) is often hindered by the prohibitive cost of repeated three-dimensional (3D) full-wave simulations, especially when the optimization is performed in a high-dimensional geometric design space. Here, we employ a computationally efficient multi-body shape optimization framework and integrate it with a wave-based linear-decoupling strategy to address this challenge. By reconstructing input-specific responses from pre-computed characteristic states, our approach bypasses the need for exhaustive, repetitive full-wave solutions of Maxwell’s equations inside the optimization loop. We utilize elliptical scatterers to enable anisotropic control over optical fields, reducing the nominal design-variable count by hundreds of times for the studied devices compared to an equivalent 40-nm pixel-based topology-optimization baseline. 3D finite-difference time-domain simulations on the Iris and MNIST classification benchmarks validate that the proposed framework yields compact ONN classifiers with high inference accuracy and significantly lower training cost. This work establishes a scalable, fabrication-aware methodology for high-performance on-chip photonic neural processors.