We report direct-write and rewritable photonic circuits on a low-loss phase-change material (PCM) thin film. Complete end-to-end PICs are directly laser-written in one step without additional fabrication processes, and any part of the circuit can be erased and rewritten, facilitating rapid design modification. We demonstrate the versatility of this technique for diverse applications, including an optical interconnect fabric for reconfigurable networking, a photonic crossbar array for optical computing, and a tunable optical filter for optical signal processing. By combining the programmability of the direct laser writing technique with PCM, our technique unlocks opportunities for programmable photonic networking, computing, and signal processing. Moreover, the rewritable photonic circuits enable rapid prototyping and testing in a convenient and cost-efficient manner, eliminate the need for nanofabrication facilities, and thus promote the proliferation of photonics research and education to a broader community.

On-chip photonic-neural-network processors have potential benefits in both speed and energy efficiency but have not yet reached the scale to compete with electronic processors. The dominant paradigm is to build integrated-photonic processors using relatively bulky discrete components connected by single-mode 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. In this talk, I will present our recent work [1] on experimentally realizing this approach with a device whose refractive index as a function of space, n(x,z), can be rapidly reprogrammed. This device combines photoconductive gain with the electro-optic effect in a lithium niobate slab waveguide. Using this device, we performed neural-network inference with up to 49-dimensional input vectors in a single pass. [1]: T. Onodera*, M.M. Stein*, et al. arXiv:2402.17750 (2024)

In this talk, examples of photonic deep neural networks for sub-nano second image classification are presented.

Superconducting optoelectronic networks are a promising route to large-scale neuromorphic computing. Using light for communication at the single-photon level overcomes fan-out challenges of electrical communication while achieving the minimum possible latency and lowest possible light levels. Such communication can be achieved with synapses based on superconducting single-photon detectors. Performing computations with superconducting Josephson junctions offers simple instantiations of synaptic, dendritic, and neuronal functions, and Josephson junctions have the highest speed over energy quotient of any known active circuit element. There is no known method to compute faster with less energy. This talk will summarize experimental demonstrations of single-photon communication links in silicon photonics, multiplanar waveguide routing networks, single-photon synapses and their associated memory circuits, optoelectronic dendrites, and superconductor-semiconductor interfaces that drive the light sources. Training algorithms that make such networks useful for artificial intelligence applications will be presented.

This talk focuses on a universal unitary architecture based on layers of diagonal matrices and specific interlacing unitary matrices. Numerical evidence shows that such an interlacing matrix can have a random pattern as long as it fulfills the required density properties. In this regard, Haar-generated random matrices are suitable candidates that lead to suitable interlacing matrices. The efficacy of such random matrices is classified by a density criterion, which provides a priori conditions to ensure the validity of the interlacing matrix in question, reducing the computational workload due to heavy numerical optimization. This approach allows for more possibilities for further interlacing matrices beyond particular waveguide arrays. Also, the random nature of the interlacing matrix provides resilience due to manufacturing defects. These results are illustrated through some full-wave simulations.

I will discuss diffractive optical networks designed by deep learning to all-optically implement various complex functions as the input light diffracts through spatially-engineered surfaces. These diffractive processors designed by deep learning have various applications, e.g., all-optical image analysis, feature detection, object classification, computational imaging and seeing through diffusers, also enabling task-specific camera designs and new optical components for spatial, spectral and temporal beam shaping and spatially-controlled wavelength division multiplexing. These deep learning-designed diffractive systems can broadly impact (1) all-optical statistical inference engines, (2) computational camera and microscope designs and (3) inverse design of optical systems that are task-specific. In this talk, I will give examples of each group, enabling transformative capabilities for various applications of interest in e.g., autonomous systems, defense/security, telecommunications as well as biomedical imaging and sensing.

Photonic computing has recently attracted strong renewed interest through experimental demonstrations in integrated photonic circuits with remarkable energy-efficiency and throughput. However, photonics alone cannot achieve general-purpose data-centric computing with agility due to fundamental limitations of bosons. We discuss nanoscale 3D-electronic-photonic-integrated-circuits (3D EPICs) and their applications in future computing. The talk will be in two parts. (a) 3D EPIC neuromorphic computing systems consisting of attojoule nanoscale optoelectronic neurons and photonic synaptic neural networks to pursue scalable brain-derived computing with self-learning capability, and (b) 3D EPIC fine-grain accelerators and memory units interconnected by all-to-all photonic networks for scalable graph analytic workloads. We will further explore heterogeneous reconfigurable computing systems enabled by photonic and electronic reconfiguration in clusters with heterogeneous 3D EPICs.

Physical Reservoir Computing is a recently developed form of Neuromorphic Computing where a non-linear physical system transfers data to higher dimensions. This information is used to train linear weights digitally to perform a specific task. Although various physical hardware have shown promising results for reservoir computing, it is not certain how to build a suitable reservoir given a set of physical components to perform a specific task. In this work, we study signal dynamics and information processing in a hybrid photonic electrical reservoir and how it can be used to build an optimal reservoir.

I will overview our recent work on training controllable physical systems to perform machine learning calculations using the backpropagation algorithm. This has the potential to enable machines that are "learned" similar to neural networks, which may be useful for computational sensing or energy-efficient machine learning. Works referenced: Wright, L. G., Onodera, T., Stein, M. M., Wang, T., Schachter, D. T., Hu, Z., & McMahon, P. L. (2022). Deep physical neural networks trained with backpropagation. Nature, 601(7894), 549-555. Senanian, A., Prabhu, S., Kremenetski, V., Roy, S., Cao, Y., Kline, J., Onodera, T., Wright, L.G., Wu, X., Fatemi, V. and McMahon, P.L., 2023. Microwave signal processing using an analog quantum reservoir computer. arXiv preprint arXiv:2312.16166. Wang, T., Sohoni, M. M., Wright, L. G., Stein, M. M., Ma, S. Y., Onodera, T., ... & McMahon, P. L. (2023). Image sensing with multilayer nonlinear optical neural networks. Nature Photonics, 17(5), 408-415.

We show that a lithographically designable network of waveguides etched into a wafer-bonded layer of InP can act as a physical neuromorphic computing system. The network is a lasing medium, with many spatially complex and overlapping modes competing for gain. The complex nonlinear interaction between these modes enables spectrally multiplexed feature detection in image data optically projected onto the network, with different spectral regions corresponding to different features. We use the network’s complex photonic dynamics to perform image classification tasks by training a single regression layer on the network’s spectral output, and construct a photonic convolutional neural network by combining the feature detection and classification layers, with 98.4% and 89.9% accuracies for MNIST and Fashion MNIST respectively. Finally, we explore how the graph properties of the network and tuning illumination parameters impact the machine learning performance of the system.

The chip industry grapples with a chip shortage and a $75 billion influx of counterfeit chips, risking malfunctions and surveillance. Several techniques to authenticate semiconductors have been introduced to detect counterfeit chips, including using physical security tags integrated into chip functionality or packaging. Among them, physical unclonable functions (PUFs), are known as unique and challenging to replicate. However, some PUFs face verification robustness challenges. We introduce a statistical and fabrication method for semiconductor device packaging that is resilient to adversarial tampering. Our innovative deep-learning approach uses a residual, attention-based discriminator to identify tampering in an optical anti-counterfeit PUF, with a random array of gold nanoparticles embedded in the package. Authentication is swiftly achieved in 80ms with 97.6% accuracy, even in challenging adversarial tampering conditions, and our approach demonstrates substantial improvements in total accuracy compared to state-of-the-art metrics. We also propose to map our concept onto a photonic neuromorphic preprocessor. Such a transition offers significant speedup and additional security.

We propose a mode-division multiplexing (MDM) approach for small-scale optical matrix multiplication and a multi-dimensional architecture for large-scale applications, supplementing wavelength-division multiplexing (WDM) with MDM to increase channel capacity for computations. We design and experimentally demonstrate key components for the MDM computing system, i.e., a multimode beam splitter, a micro-ring modulator for high-order modes, and a multimode waveguide bend. We further show a 2-by-2 matrix multiplexing system fabricated in a foundry that works for both MDM and MDM-WDM computing, which confirms that our approach successfully increases the input vector size for computing and ensures compatibility with existing WDM networks.

Neuromorphic computing seeks to emulate the brain's efficiency in hardware, enabling sophisticated information processing with minimal energy consumption. Despite these advances, dynamically training neuromorphic systems—whether electronic or photonic—to undertake new tasks directly within the hardware remains a formidable challenge. This is largely due to the nature of conventional training algorithms like backpropagation. They are not well-suited for hardware implementation because they rely on non-local computations. In this presentation, I will introduce a solution to this challenge through the application of Equilibrium Propagation, an algorithm rooted in physics, first proposed by Scellier and Bengio in 2017. This innovative approach leverages the inherent dynamics of physical systems for learning, capitalizing on how these systems can be gently guided ('nudged') towards a desired outcome through their natural behavior. I will elucidate the foundational principles of Equilibrium Propagation, discuss the obstacles encountered in its implementation, and demonstrate its versatility through applications across a variety of physical systems.

The maximum independent set (MIS) problem is an optimization problem in graph theory and is related to various applications such as network design and resource allocation. Here we report that a coherent Ising machine based on a network of optical parametric oscillators can find large independent sets in graphs that consist of thousands of nodes faster than simulated annealing implemented many-core CPU.

This paper proposed to adopt advanced monolithic silicon-photonics integrated-circuits manufacturing capabilities to achieve a system-on-chip photonic-electronic linear-algebra accelerator with the features of broadband incoherent photo-detections and high-dimensional operations of consecutive matrix-matrix multiplications for enabling substantial leaps in computation density and energy efficiency with practical considerations of power/area overhead due to photonic-electronic on-chip conversions, integrations and calibrations through holistic co-design approaches to support attention-head mechanism based deep-learning neural networks used in Large Language Models and other emergent applications.

We demonstrate how we can use co-designed meta-optics and computational backend to perform object recognition. Specifically we analyzed the regime, where we observed a distinct advantage of the hybrid approach over a pure digital one.

Specialized hardware accelerators have been proposed to improve the throughput and energy-efficiency of deep neural network (DNN) models. However, collective data movement primitives such as multicast and broadcast that are required for multiply-and-accumulate (MAC) computation in DNN models are expensive and require excessive energy and latency when implemented with electrical networks. Emerging technology such as silicon photonics can inherently provide efficient implementation of multicast and broadcast operations, making photonics more amenable to exploit parallelism within DNN models. Moreover, when coupled with other unique features such as low energy consumption, high channel capacity, silicon photonics could potentially provide a viable technology for scaling DNN acceleration. In this talk, I will discuss an analog photonic architecture for scaling DNN acceleration using microring resonators and Mach-Zehnder modulators. Using detailed device models, an efficient broadcast combined with multicast data distribution by leveraging parameter sharing through unique WDM dot product processing will be discussed.

We propose RAPT, a reconfigurable accelerator for photonic training. RAPT is a low-power photonic accelerator that combines the benefits of phase change material (PCM) and photonics to implement both inference and training in one unified architecture, reconfigurable to support a variety of neural network models. Emerging silicon photonics has the potential to exploit parallelism of neural network models, reduce power consumption and provide high bandwidth density via wavelength division multiplexing making photonics an ideal candidate for on-device training and inference. As PCM is reconfigurable and non-volatile, we utilize PCM for the unique purposes of maintaining resonant wavelength without expensive electrical or thermal heaters and implementing non-linear activation function which eliminates the need to move data between memory and compute units; both of which leads to significant reduction in energy consumption and execution time.

Photonic neural networks (PNNs) have been attracting intense attention owing to their low latency and low-power consumption. Among the PNNs, the optoelectronic recurrent neural network (RNN) enables low-power and high-speed time-series data processing using a compact loop structure. However, the RNN performance can be limited by the resistive-capacitive (RC) delay of optical-electrical-optical (OEO) converters. Here, we study in simulation an optoelectronic RNN equipped with OEO converters with RC delay. We confirm the correct RNN behavior even when RC delay is comparably large to the time interval of data. Our simulation results show that the optoelectronic RNN achieves 80% test accuracy for the sequential MNIST task and that the accuracy increases with RC delay when loop losses exist. Our analyses reveal that the accumulation of time-series data by RC delay does not degrade the RNN performance but rather can compensate for the degraded RNN performance due to loop losses.

To scale up programmable photonic circuits, one must overcome challenges of size, power, and robustness to hardware imperfections. Considering linear multiport interferometers as a concrete example, we show how architectural choices can significantly improve the prospects for scalability. First, we review the field of photonic-circuit error correction, and show how 3-splitter MZI (3-MZI) architectures are more robust to errors than standard MZI architectures, even achieving asymptotic fault-tolerance to hardware imperfections. Second, we discuss the problem of economizing phase-shift, which is especially relevant for power-hungry thermo-optic platforms, and show how the 3-MZI architecture can reduce average phase shifts by a factor of 3-10x in near-term systems.

Traditional neural networks (NNs) require substantial computing resources and energy. Photonics-based systems offer faster and more energy-efficient solutions. However, implementing nonlinear activation functions in photonics has been challenging due to the need for high-power optical sources and extended interaction lengths. The proposed solution uses structural nonlinearity, creating nonlinear output patterns with low energy use and simple digital NN training. The research develops a reconfigurable material platform using liquid crystal/polymer composite (LCPC) and metasurface to control scattering potentials dynamically. These results show that the LCPC’s phase distribution can be reliably controlled, enabling reconfiguration and repetition of scattering responses, which is crucial for advancing photonics-based neuromorphic computing.

The recent advancement of photonic neural networks (PNNs) enables high-throughput, low-power neuromorphic computing, bypassing capacitive wire limitations. Here, we discuss advanced approaches to modeling an optical perceptron based on reverse saturated absorption (RSA). To enable real-life models beyond classical Lorentz and Debye media, we derive the novel frequency dispersion equations of RSA materials with disorder and couple them with the master equations for singlet and triplet state carrier kinetics in a joint finite difference time domain framework. Our numerical results match the experimental measurements; the proposed framework can enable efficient optimization of all-optical RSA perceptrons for next-generation photonic hardware.

Among the emerging approaches for deep neural network (DNN) acceleration, analog in-memory computing (AIMC) emerges as a promising solution by directly performing computations within crossbar arrays of non-volatile memories (NVMs), using the device conductance values to represent the weight. Phase change memory (PCM) is one of the most promising NVMs for AIMC particularly for inference using previously trained DNN weights. However, the utilization of PCMs for analog computing introduces nonidealities, including resistance drift, read noise, limited memory window, and various device failures. I will discuss our work on optimizing PCM devices to alleviate these nonidealities and mitigate their impact on AIMC.

Here we show efficient ways of optical vortex beams generations by using spherulite materials.

Photonic solvers capable of finding the ground states of various spin Hamiltonians can be employed to investigate numerous interactive physical systems and combinatorial optimization problems. We introduce a novel photonic solver that leverages spin Hamiltonians within the framework of Modulated Momenta of Light. We establish a real-and-momentum space correspondence for spin Hamiltonians using spatial light transport. The real-space spin interaction is determined by modulating the momentum-space flow of light. This principle, formulated as a generalized Plancherel theorem, enables us to construct a straightforward optical simulator capable of finding ground states for displacement-dependent spin interactions. In our experiments, we apply this principle to unveil the exotic magnetic phase diagram of iron chalcogenides, a family of iron-based superconductors, and we also observe vortex-mediated Berezinskii-Kosterlitz-Thouless dynamics. The precision achieved by momentum modulation in controlling spin interactions holds promise for exploring novel physical effects and solving complex optimization problems.