Speaker
Description
Diffractive neural networks (DNNs) exploit the nature of light propagation in free space and the interaction of a light field with metasurfaces designed with machine learning (ML) methods to implement inference passively in the optical domain1. In this way, DNNs can achieve all-optical ML tasks such as image classification, image compression or decryption2. When fabricated with two photon nanolithography (TPN) methods, DNNs can express their full potential achieving a neuron density of >500 million neurons per square centimeter, operative wavelength in the near infrared or visible wavelength regime and on-chip integration with imaging sensors, such as commercial complementary metal-oxide-semiconductor (CMOS) sensors3. This co-integration is particularly relevant because fully passive nanoprinted DNNs also have intrinsic limitations. Their performance can be highly sensitive to printing errors, alignment tolerances, optical aberrations, and deviations between the simulated and fabricated structures4. After fabrication, the diffractive elements are static and cannot easily be reconfigured or tuned. Moreover, the implementation of hidden optical nonlinearities remains challenging, although nonlinearity is essential for solving more complex ML tasks. Additional limitations arise from the training process itself: many current models rely on approximate scalar diffraction physics, simplified material descriptions, and limited treatments of nanophotonic effects, fabrication constraints, and energy efficiency. Hybrid optoelectronic architectures offer a practical route to overcome part of these limitations. In these systems, the nanoprinted DNN performs large-scale optical transformations and information compression, while the CMOS sensor provides optoelectronic conversion and a nonlinear digital interface that can be modelled, in first approximation, as a shifted ReLU activation. The digitized output can then be processed by a compact electronic neural network. For selected tasks, such hybrid optoelectronic neural networks can reach performances comparable to purely digital networks while substantially reducing the computational load required in the digital domain. In this talk, I will discuss both the potential and the current bottlenecks of nanoprinted DNNs, including fabrication-error sensitivity, static operation, nonlinear activation, training strategies, approximate physical modelling, limited nanophotonic design frameworks, and alignment constraints. I will then present several directions from our work aimed at addressing these challenges, including the development of energy-aware hierarchical optimization strategies for DNNs, the study of optical nonlinearity as a physical computational resource and the exploration of analog photonic computing with engineered materials. Together, these approaches point toward a broader framework in which nanoprinted optical neural networks are not treated simply as passive phase masks, but as physically engineered photonic computing platforms.
References
1. Lin, X. et al. All-optical machine learning using diffractive deep neural networks. Science 361, 1004–1008 (2018).
2. Goi, E. et al. Nanoprinted high-neuron-density optical linear perceptrons performing near-infrared inference on a CMOS chip. Light Sci. Appl. 10, 40 (2021).
3. Goi, E., Schoenhardt, S. & Gu, M. Direct retrieval of Zernike-based pupil functions using integrated diffractive deep neural networks. Nat. Commun. 13, 7531 (2022).
4. Chen, M., Schoenhardt, S., Gu, M. & Goi, E. Quantitative comparison of the computational complexity of optical, digital and hybrid neural network architectures for image classification tasks. Opt Express 31, 44474–44485 (2023).