发表机构
ITMO University; University of Brescia（圣彼得堡信息科学技术大学; 布雷西亚大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究提出用单个米氏散射体实现神经形态光学计算，克服传统衍射处理器参数密度限制，在ka=15时对相位编码MNIST图像分类准确率达90%，可实现片上集成。
AI 中文摘要
光学计算正成为下一代信息处理的极具前景的范式。衍射光学处理器依赖空间分布的可训练自由度，导致架构扩展。在此，我们提出一种紧凑的神经形态光学计算方法，其中整个可训练变换由单个米氏散射体实现。通过在矢量球谐基中构建计算，可训练的模式耦合可通过散射体的T矩阵集中在有限物体内。由于可用T矩阵参数的数量与最大多极阶数的四次方成比例，该架构可克服传统空间分布衍射处理器的可训练参数密度限制。我们利用散射场强度演示了相位编码MNIST图像的分类。在粒子尺寸参数ka=15时，训练后的T矩阵达到约90%的测试准确率，与单层人工神经网络相当。利用近场也可实现类似性能，支持片上集成。我们进一步证明互易性、无源性和粒子对称性如何约束性能：无源性减小了可访问的算子空间，但提高了鲁棒性；对称性减少了独立参数的数量。最后，我们逆向设计了一种非吸收性电介质散射体，其分类任务准确率达84%。这些结果表明，非平凡的神经形态变换可编码在单个紧凑散射体的多极响应中。
英文摘要
Optical computing is emerging as a promising paradigm for next-generation information processing. Diffractive optical processors rely on spatially distributed trainable degrees of freedom, leading to extended architectures. Here, we propose a compact neuromorphic optical-computing approach where the entire trainable transformation is implemented by a single Mie scatterer. By formulating computation in vector spherical harmonics basis, trainable modal couplings can be concentrated within a finite object through its T-matrix. Since available T-matrix parameters scale as the fourth power of maximal multipole order, this architecture can overcome trainable-parameter-density limitations of conventional spatially distributed diffractive processors. We demonstrate classification of phase-encoded MNIST images using scattered-field intensity. At a particle size parameter of $ka = 15$, the trained T-matrix reaches approximately 90% test accuracy, comparable to a single-layer artificial neural network. Similar performance can be achieved using near-fields, enabling on-chip integration. We show how reciprocity, passivity, and particle symmetry constrain performance: passivity reduces the accessible operator space while improving robustness, whereas symmetry reduces the number of independent parameters. Finally, we inverse-design a non-absorbing dielectric scatterer that realizes the classification task with 84% accuracy. These results demonstrate that nontrivial neuromorphic transformations can be encoded within the multipolar response of a single compact scatterer.
