Scaling on-chip photonic neural processors using arbitrarily programmable wave propagation

TL;DR

The paper tackles the scalability bottleneck of on-chip photonic neural networks by moving from networks of discrete components to a continuously programmable 2D waveguide within a LiNbO3 slab. It demonstrates a four-layer device whose refractive-index distribution n(x,z) is reprogrammable across ~10^4 degrees of freedom via photoconductive gain and the electro-optic effect, enabling massively parallel wave propagation. The authors show single-pass neural inference on vowel and MNIST tasks with 96% and 86% accuracy, respectively, highlighting the potential of continuous-waves photonics to achieve higher integration density and compactness than conventional photonic chips. This work paves the way for compact, versatile on-chip photonic systems for optical processing, sensing, spectroscopy, and communications, with clear paths toward deeper networks, higher throughput, and broader functionality through materials and integration advances.

Abstract

On-chip photonic processors for neural networks have potential benefits in both speed and energy efficiency but have not yet reached the scale at which they can outperform electronic processors. The dominant paradigm for designing on-chip photonics is to make networks of relatively bulky discrete components connected by one-dimensional waveguides. A far more compact alternative is to avoid explicitly defining any components and instead sculpt the continuous substrate of the photonic processor to directly perform the computation using waves freely propagating in two dimensions. We propose and demonstrate a device whose refractive index as a function of space, $n(x,z)$, can be rapidly reprogrammed, allowing arbitrary control over the wave propagation in the device. Our device, a 2D-programmable waveguide, combines photoconductive gain with the electro-optic effect to achieve massively parallel modulation of the refractive index of a slab waveguide, with an index modulation depth of $10^{-3}$ and approximately $10^4$ programmable degrees of freedom. We used a prototype device with a functional area of $12\,\text{mm}^2$ to perform neural-network inference with up to 49-dimensional input vectors in a single pass, achieving 96% accuracy on vowel classification and 86% accuracy on $7 \times 7$-pixel MNIST handwritten-digit classification. This is a scale beyond that of previous photonic chips relying on discrete components, illustrating the benefit of the continuous-waves paradigm. In principle, with large enough chip area, the reprogrammability of the device's refractive index distribution enables the reconfigurable realization of any passive, linear photonic circuit or device. This promises the development of more compact and versatile photonic systems for a wide range of applications, including optical processing, smart sensing, spectroscopy, and optical communications.

Figures (24)

• Figure 1: Machine learning with multimode wave propagation in a 2D-programmable waveguide.A, The fundamental unit of an artificial neural network---a layer---transforms an input vector into another output vector via a trainable matrix multiplication. B, Analogous to a neural network layer, the 2D-programmable waveguide linearly transforms an input optical field into an output optical field, via wave propagation through a slab waveguide (red slab in the device stack) whose refractive index distribution can be continuously and arbitrarily programmed. This refractive index distribution, which is directly set by the illumination pattern (shown in green) that is projected onto the device---whose top layer is a photoconductive film (yellow), is trained to perform machine-learning tasks. To determine the result of the computation, we measure the output beam's intensity across equally-sized bins; the bin with the highest total power corresponds to the predicted classification label. C, Simulated intensity distribution of the optical field within the slab waveguide, after training of the parameters. It illustrates that the neural-network computation is performed with complex multimode wave propagation.
• Figure 2: Operating principle of the 2D-programmable waveguide.A, The 2D-programmable waveguide consists of a nanophotonic stack of four layers: 1) A conductive silicon substrate that doubles as the ground electrode, 2) a Z-cut lithium niobate (in red) slab waveguide with silicon dioxide cladding (in white), 3) a photoconductive layer for optical control of the refractive index, and 4) a gold electrode. B-C, Electrical circuit models of the 2D-programmable waveguide in regions with and without illumination. There is a voltage division between the photoconductor and the lithium niobate slab waveguide, with impedances $Z_{\text{pc}}$ and $Z_{\text{wg}}$, respectively. C: Upon illumination, the resistance of the photoconductor decreases, leading to an increase in the electric field (illustrated with blue arrows) inside the waveguide. This induces a refractive index modulation in illuminated regions via the strong electro-optic effect in lithium niobate. D, A photograph of the 2D-programmable waveguide in the experimental setup. E, Experimental realization of a Y-branch splitter on the 2D-programmable waveguide, which splits the input light into two equal output beams. The projected pattern (in green) directly corresponds to the refractive index distribution (in gray), subject to spatial smoothing and a weak nonlinearity due to saturation of the photoconductor. A simulation of the wave propagating through the refractive index pattern is overlaid with the patterns (in red).
• Figure 3: Vowel classification with the 2D-programmable waveguide.A, Overview of approach: The task involves predicting a spoken vowel, here "er", from a 12-dimensional input vector representing formant frequencies extracted from audio recordings. The 2D-programmable waveguide is trained to perform computation on this input vector, producing a 7-dimensional output vector where the vector index with the highest value corresponds to the predicted vowel. B, Left: The input vector is amplitude-encoded into twelve Gaussian spatial modes to produce the initial optical field distribution. Center: Simulated wave propagation in the chip after training of the projected pattern. Right: The experimentally measured output intensity. It is binned, i.e., the total power within equally-spaced spatial bins is calculated to produce the 7-dimensional output vector. C, Illustration of physics-aware training, a hybrid in-situ, in-silico backpropagation algorithm, which we use to train the parameters of the 2D-programmable waveguide. The forward pass is performed by the experimental setup, while the backward pass is computed with a digital model of the experiment. D, Test accuracy as a function of epoch. The inset shows the confusion matrix for the test dataset of 63 vowels, after 300 epochs of training. E, Evolution of the trainable parameters, the projected patterns, at different stages of the training.
• Figure 4: Neural-network inference with a high-dimensional input vector: MNIST handwritten-digit classification.A, We perform MNIST handwritten digit classification with the 2D-programmable waveguide. Images from the MNIST dataset are electronically downsampled and reshaped to a 49-dimensional vector. We train the device to perform machine learning on this high-dimensional input vector with the same procedure as the vowel classification task (see Fig. 3). B, The confusion matrix for the test dataset of 10,000 images. After 10 epochs of training, the system achieved 86% accuracy on the test dataset. As a baseline, a single layer digital neural network with a $49\times10$ matrix achieves 90% accuracy on this same task.
• Figure A1: Schematic of the fabrication process for the 2D-programmable waveguide.