Data-driven machine learning framework has become a state-of-art technique to explore whole parameters design space for designing complex systems. In this work, we used conditional generative adversarial networks to inverse design three problems that we are interested in random nanophotonic systems: pattern optimization, geometry generation, and pattern reproduction. Meanwhile, automation convolutional neural networks group for forward prediction of the transmission spectra of disordered waveguides in linear and nonlinear regimes, at telecommunication wavelengths.

The quantitative evaluation of plant organs in a non-destructive and continuous fashion is the technological bottleneck to meet the food, fuel, and fiber needs for the 10 billion people on earth by 2050. Quantifying crop root architecture paves promising ways to improve resource uptake in the face of the resource limitations in the degraded soils of future climates. Current root measurement methods either have low resolution or involve uprooting the plant. In all cases, the measurement methods do not provide any prediction on how well the plant is growing. We propose the usage of three fiber Bragg gratings (FBG) embedded within soil to measure underground strain change due to pseudo-root growth and a Residual Neural Network (ResNet) to predict its characteristics in a non-destructive fashion. To generate large amounts of sensor data similar to that of a growing root, we developed an automated robot that inserts pseudo-roots of 1mm and 5mm in diameter to 15cm below the soil’s surface over the span of 11 minutes. We used 2,582 and 240 samples in training of the diameter and depth models, while testing was performed using 646 and 60 samples. The models were able to achieve accuracy of 92% and 93% for diameter and depth prediction, respectively. Through transfer learning, our base models will be expanded so that real time prediction on actual plant roots diameter and depth can be achieved.

This work proposed a universal platform for ultra-sensitive detection, which integrates sensory data acquisition and spectral feature extraction into a single machine learning (ML) hardware.We fabricated and tested the sensing platform in glucose detection tasks, reaching 5 orders of magnitude higher sensitivity compared to the state-of-the-art. This technology requires no bulky spectral measuring devices such as a spectrum analyzer but a standard off-the-shelf camera to achieve real-time detection of the glucose concentration.

We propose a new feedback correction system driven by artificial intelligence (AI), in particular reinforcement learning (RL), able to learn from the turbulence pattern how to correct the distortions. Indeed, RL is utilized to solve difficult tasks in chaotic problems making predictions based on the environment responses. We apply this novel approach in a Quantum Key Distribution (QKD) free space horizontal link field-trial test within the metropolitan area of Florence operating the Quantum Communication in the third telecommunication window (1550nm) with time-bins states. We use the combination of a fast-steering mirror (FSM), a four-quadrant detector (QD), and a closed-loop to correct the turbulence-induced beam-wandering effect. Our closed-loop architecture is composed of a core Proportion-Integrative-Derivative (PID) controller and an auxiliary RL algorithm to find the optimal P, I, and D parameters. We demonstrate the robustness and effectiveness of using the RL approach to smooth the turbulence effects in communication.

In this talk, we report resonance tuning of a silicon nitride microring resonator structure using photochromic molecules. A slot waveguide structure and back-end compatible light molecule evaporation are used to enhance interaction of the molecules and optical mode. The device is interrogated in the optical C-band where the molecules exhibit low optical loss, but where a change in refractive index is present. Under UV illumination the resonance is observed to redshift, while under visible illumination the resonance blueshifts. Furthermore, the observed index shift is seen to be non-volatile. This constitutes a new way to optically reversibly trim and reconfigure high index contrast photonic integrated circuits for which a plethora of applications have been investigated recently.

We use a genetic algorithm to optimize periodic arrays of truncated square-based pyramids made of alternating stacks of metal/dielectric layers. The objective is to achieve broadband quasi-perfect absorption of normally incident radiations in the visible and near-infrared ranges (wavelengths comprised between 420 and 1600nm). We compare the results one can obtain by considering one, two or three stacks of (i) Ni, Ti, Al or Cu for the metal, and (ii) poly(methyl methacrylate) (PMMA) for the dielectric. The parameters to determine for each metal/dielectric combination are (i) the period of the system, (ii) the lateral dimensions of each stack of metal/dielectric layers and (iii) the width of each dielectric layer. The Rigorous CoupledWaves Analysis (RCWA) is used to compute the absorptance spectrum of these different structures. The Genetic Algorithm is used to find the geometrical parameters that maximize the integrated absorptance. This approach provides stability maps with respect to the geometrical parameters, which leads to additional physical insight regarding practical implementation. The study shows that Ni/PMMA and Ti/PMMA provide high-quality solutions associated with broad optima (stability with respect to variations of the geometrical parameters). On the contrary, Al/PMMA and Cu/PMMA provide poor-quality solutions associated with sharp optima. We find an interesting correlation between the robustness of the solutions found by the genetic algorithm (stability with respect to variations of the geometrical parameters) and the stability of these solutions with respect to the number of plane waves used in the RCWA calculations.

Topology optimization was used to design various optical and photonic solid-state devices. The designs of those devices are commonly composed of only two materials with different refractive indexes, which means that the refractive index is not continuously spatially varying. With additive manufacturing, photoalignment and similar techniques it is possible to make almost arbitrary designs of soft-matter photonic devices. The advantage of such devices over solid-state devices is that the refractive index can continuously change, which can improve performance, and soft-matter devices can be more cost-effective to manufacture. We use topology optimization in combination with a FDTD solver to design soft-matter diffraction gratings for linearly polarized light with the first diffraction order at a specific angle. During the optimization process, we consider material and manufacturing constraints, such as structure relaxation and maximum feature sizes due to the elastic energy associated with the designed structure and chosen material. The diffraction gratings are optimized for light in the IR and visible parts of the spectrum. We calculated designs of soft-matter diffraction gratings which can be manufactured using photopatterning or additive manufacturing and diffract light at the designed angle for specific wavelengths. By using topology optimization for soft-matter optical/photonic devices we can improve their quality and, in some cases, create low-cost alternatives to solid-state devices.

In conventional metasurface structural colour design, simulations combined with human intuition are used for design and optimization, making it challenging to find the best solution. Here we introduce an innovative AI-assisted design process that bypasses the need for complex simulations, enabling swift and precise mapping between metasurface parameters and colour coordinates. Instead of assigning one colour to one geometry, we demonstrate that multiple colours can be generated from a single geometry under varying levels of strain. This can be achieved through a single model, facilitating the development of active metasurfaces using AI. This finding enables designers to create active metasurfaces that account for both geometric properties and dynamic responses in a unified model which could accelerate the development of active metamaterials closer to practical applications in the real world.

Silicon nanostructures have a rich optical response thanks to Mie-type optical resonances, that can be designed on-demand via their geometry. It is possible to encode bits of information in a nanostructure’s geometry, and retrieve this information optically via the color observed in dark-field microscopy. Furthermore, asymmetric structures can profit from the illuminating light polarization to facilitate information readout. Our ultimate goal is to accurately reverse engineer experimentally feasible silicon nanostructures for information encoding, such that they implement a set of ideally distinguishable colors for robust optical readout. Deep learning is increasingly being used to solve inverse problems such as nano-photonic structure design. Neural networks for inverse design are mostly trained on simulated data, which is cheap to generate. But training neural networks on experimental data is a very interesting option, because it allows to include all experimental constraints into the model, which consequently learns to capture phenomena that may be hard to simulate. Here, in order to learn an accurate model for the full experimental measurement setup, we trained a neural network with experimental darkfield color data from several thousand nanostructures. Firstly, we built a forward network, taking as input the nanostructures’ shapes from fabricated samples and predicting the dark-field color for both X and Y polarizations. We then successfully built an inverse tandem network, capable of designing structures with desired color responses. In order to create distinguishable color responses, another deep neural network was trained on the task to map all experimental colors in a regularized color latent space. Sampling equidistant points from this latent space then yields the most distinguishable, yet experimentally feasible colors. The next future step will be to produce samples from the generated structures to test the network’s accuracy. We would like to test how many bits of information we can encode using the darkfield color as readout.

The field of metamaterials allows for the creation of materials with extraordinary properties. However, the development of materials with specific, custom properties is regarded as a challenging task. Current materials-by- design methodologies hinge on a trial-and-error approach, employing serendipitous techniques that prove inefficient and impractical. Furthermore, the extensive variety of materials and the myriad ways they can be combined in different ratios contribute to an infinite compositional space. Here, we present a universal machine-learning method that identifies the complex, nonlinear relationship between an amorphous metamaterial’s structural characteristics and on-demand optical properties, all within a matter of milliseconds. As a proof of concept, we have demonstrated two practical applications of the method experimentally by developing a custom metasurface-perfect reflector. This innovative approach empowers users to craft materials of interest without depending on intuitions, prior experiences, or extensive simulation and modelling, potentially paving the way for the accelerated discovery of new materials.

Structural colour filters can display various colours by selectively transmitting or reflecting a specific wavelength by varying structural parameters rather than material components in the visible region. An important aspect of structural colour is the ability to design a structure that can accurately display the desired colour. While the conventional trial-anderror method requires substantial prior knowledge of the structure together with a number of simulations, deep learning provides an alternative way to inverse design the structural colour with high efficiency and accuracy. In this abstract, we will be discussing the deep learning enabled inverse design of structural colour. By employing the conditional generative adversarial networks (cGAN) to inverse design the structural colour, the one-to-many problem that is often encountered in nanophotonic inverse design is fully tackled. Moreover, we will also explore the possibility of applying this system to the dynamic structural colour inverse design.

Machine learning techniques have been proposed in the literature for the modeling of photonic devices. These techniques can be used to speed up the design process. The data samples needed to build machine learning models are collected from electromagnetic simulations. Electromagnetic solvers can result computationally expensive and therefore minimizing the computational effort needed to collect these data samples is an important aspect. Using frequency-domain electromagnetic solvers to collect data samples requires a suitable sampling of the wavelength variable to avoid undersampling and oversampling phenomena. An adaptive frequency-domain sampling approach for nanophotonic applications is illustrated in this work.

The present work showcases an innovative optimization methodology based on deep learning that combines Multi- Valued Artificial Neural Networks and back-propagation optimization. The methodology addresses the inherent limitations of conventional approaches when employed in isolation. We applied the proposed methodology to design structural color filters that surpasses the sRGB gamut while preserving fabrication constraints.

In this work, we propose a novel framework for large-scale aperiodic nanophotonic inverse design utilizing an experimental machine-learning technique. With this technique, we create an extensive dataset of 10 million experimental structures for enhanced flat-optics design. This largest publicly available inverse design dataset, achieved through electron beam lithography, bypasses the extensive computational demand of first-principle simulations. Experimental acquisition ensures the dataset embodies real-world variances, leading to ML models with a ten-fold improved prediction accuracy in optical responses, drastically reducing validation RMSE from 0.012 to 0.0018. With this dataset, we developed a framework for large-scale aperiodic photonics design capable of designing tens of structures per second. We demonstrate the efficiency of the proposed technique by creating a large (3x3mm) aperiodic photonic structure composed of >10000 individual structures with pre-defined transmission/reflection responses.

The adjoint method is an efficient technique for the topology optimization of complex nanophotonic systems, including nanostructures, metasurfaces and integrated optical circuits. While such method has been traditionally used in the frequency domain, its extension to the time domain opens new opportunities for wideband optimization of dispersive materials for applications ranging from broadband absorbers to enhanced quantum emitters in dispersive environments. We propose a topology optimization technique for the inverse design of linear optical materials with arbitrary dispersion and anisotropy. We introduce a general adjoint scheme in the time-domain based on the complex-conjugate pole-residue pair (CCPR) model. This approach has the advantage of treating dispersive media and broadband response naturally in a single simulation run. We implement this framework within the finite-difference time-domain (FDTD) method and investigate the method for optimizing metallic and dielectric nanoantennas over the optical spectral range of 350 to 1000nm. The combination of the method with parallel computing enables the large-scale inverse design of nanostructures in 3D with extreme field confinement. Nanostructures found via inverse design and featuring the intriguing anapole effect are also discussed. This effect enables nanostructures that show field enhancement, negligible scattering, and low losses. The possibility of reducing losses in plasmonic nanostructures via inverse design is an interesting possibility offered by the method and may open new avenues towards the realization of transparent plasmonic metamaterials for applications in linear and nonlinear nanophotonics.

In the past decade, the field of neuromorphic photonics has experienced significant growth. To extend the reach of this technology, researchers continue to push the limits of these systems with respect to network size and bandwidth. However, without proper RF-optimized architectural designs, as operating frequencies are scaled up, significant losses of RF power can be incurred at each neuron. Within the broadcast and weight neuromorphic photonic architecture, this excess loss will be accumulated until processing is no longer feasible. If designed properly, RF loss can be minimized significantly, and residual loss could be compensated by cointegrated transimpedance amplifiers, thus enabling further scaling of the network. In this paper, the authors present broadband weighting of RF input signals with a 3-dB bandwidth of 4.28 GHz, utilizing the linear front-end of a silicon photonic neural network. Additionally, the authors present link loss measurements and analysis.

Machine learning is a critical tool for sensing due to its ability to process and interpret complex sensor data, as well as to enhance the accuracy and efficiency of sensing applications in diverse fields. This paper provides an overview of machine learning’s multifaceted applications in microwave photonics, soft robotics, and precision agriculture sensing. Recently, machine learning techniques have revolutionized the field of microwave photonics. As an example, we will discuss an implementation of deep learning and generative adversarial network for data argumentation in instantaneous frequency measurement, which effectively decreases required training experimental dataset size by 98.75% and reduces error to <5%. Enhancing the practicability and accuracy of the system. Next, we shift our focus to the integration of fiber optic sensors in soft robotics to offer a lightweight, compact, and soft means of analyzing important robot parameters. By utilizing sensor data, machine learning algorithms enable real-time feedback, adaptability, and improved control of soft robot. Lastly, we also developed fiber optic sensors for non-invasive and continuous underground monitoring of root growth. Monitoring plant root growth is essential for agriculture; however, strain generated by the growth of root is relatively weak and noisy. Therefore, data collected by these fiber sensors is fed to a residual neural network to facilitate extraction of meaningful insights. In summary, machine learning has driven substantial progress in various fields that elevates the levels of accuracy and efficiency beyond previous achievements.

Graphene is a two-dimensional material with a hexagonal lattice structure made up of carbon atoms, which has exceptional optoelectronic properties such as high thermal conductivity, broad light absorption wavelength, and ultrafast carrier mobility. As a result, various graphene-based optoelectronic devices have been developed with exceptional performance. In this talk, an ultrafast all-optical nonlinear activator with a response time of 93.6 ps will be presented. This breakthrough in speed of all-optical nonlinear activators opens the door to significant improvements in the performance and applicability of ONNs.

Three-dimensional Optical Neural Networks (ONN) are a promising solution to the energy, time, and area yearning Artificial Intelligence (AI) hardware. The 3D additive manufacturing technique with Two-Photon Polymerization (TPP) can be used to build the 3D dense ONN. In our work, we designed and fabricated the hybrid waveguide circuit which fuses the polymer and air clad waveguides, an important interconnect for the ONN. The polymer-cladded waveguide can support single mode and evanescent coupling while the air-cladded can support tight bend for dense integration.

Optical neural networks enable massively parallel and energy-efficient computing, making them a promising candidate for future sustainable computing architectures. However, the non-differentiability of these systems prohibits gradient-based optimisation, making training these networks a significant challenge. We introduce a meta-learning scheme that employs reinforcement learning to generate a gradient-free optimiser capable of training physical networks on various tasks in situ. The learned optimiser can improve training time and final accuracy compared to existing gradient-free methods when training a diffractive optical network on a variety of image classification tasks, providing a new option for gradient-free training general neuromorphic systems in situ.

On-chip photonic-neural-network processors promise 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 discrete components connected by single-mode waveguides. A far more compact alternative is to avoid discrete components and instead sculpt a complex and continuous microphotonic medium in which computations are performed by multimode waves controllably propagating in two dimensions. We show our realization of this approach with a device whose refractive index as a function of space can be rapidly reprogrammed. We demonstrate optical computations much larger and more error-resilient than previous photonic chips relying on discrete components. We argue that beyond photonic-neural-network processors, devices with such arbitrarily programmable index distributions enable the realization of a wide range of photonic functionality.

Photonic machines based on spatial optics are promising for optimization and machine learning at a large scale and enable novel functionalities across photonics. We present the topic with a review of our work on three main paradigms. The first is the spatial photonic Ising machine (SPIM), which exploits spatial light modulation and coherent optical propagation to solve hard combinatorial optimization problems by taking advantage of optical parallelism and scalability. The second is the photonic extreme learning machine (PELM) based on free-space optics. We show a large-scale experimental implementation with half a million addressable nodes, which allows us to perform photonic machine learning in the so-called over-parametrized region and to implement photonic natural language processing. Finally, we exploit the computing principle of spatial photonic learning machines to demonstrate single-shot polarization imaging. This original method enables a new fast and compact polarization camera for the many contexts where conventional polarimetry is unavailable, thus opening new possibilities in imaging and optical communication.

With increasing computational demands, the importance of efficient computer architectures rises. Photonic Ising machines offer a promising approach to address complex binary optimization problems by leveraging the speed and bandwidth inherent in photonic systems. When the cost function of an optimization problem can be mapped to the energy of an Ising problem, the ground state of the latter will represent the optimal solution. We conduct a stability analysis on this ground state of various benchmark problems. This analysis reveals that some benchmark problems are manageable for all Ising machine implementations. However, there are other problems for which the probability of achieving the ground state depends on the physical implementation. Our analysis shows the reasons for this advantage, leading us to formulate strategies to further enhance Ising machines and pave the way for future research. In future work, we will investigate if photonic Ising machines utilizing a different nonlinearity possess a distinct advantage in this context.

An Ising machine (IM) is a novel natural computing system that could solve complex optimization problems more efficiently than traditional digital computers. In IMs, a problem is encoded into a network of non-linear elements, where the optimal solution is found by minimizing the network’s energy through an iterative feedback loop. Most of modern-day photonic-IM implementations have a hybrid layout where the problem variables are implemented in the analog domain, but the feedback signal is first calculated on traditional digital hardware before converted back into an analog signal. This latter conversion is implemented by an optical modulator, which have a significantly lower resolution than electronic modulators. Using numerical simulations, we investigate the influence of the limited modulator resolution on the performance of photonic-IMs and demonstrate that a 1-bit-resolution can improve the time-to-solution by at least an order of magnitude.

In the context of optical computing, photonic reservoir computing emerges as a scalable, energy-saving, and noise-robust alternative to quantum computing for machine learning. However, existing methods often lack the flexibility to finely control nonlinearities in the optical reservoir for improved performance. Here, we propose a novel photonic reservoir computing system based on spatial nonlinear wave propagation in erbium-doped multimode fibres (ED-MMF). Utilising phase-only spatial light modulators, we structure pump and probe beams in the fibre to encode and process information. Through nonlinear interactions between signal and pump modes within the gain medium, the ED-MMF enables a tunable nonlinear transformation of the input field, allowing control over nonlinear coupling between different fibre modes via accessible parameters like pump and signal power. By dynamically tuning the degree of nonlinearity, our system can identify optimal operating conditions for the reservoir, promising enhanced optical computing capabilities with potential applications in advanced machine learning tasks.

Silicon microring resonators (MRRs) have shown strong potential in acting as the nonlinear nodes of photonic reservoir computing (RC) schemes. By using nonlinearities within a silicon MRR, such as the ones caused by free-carrier dispersion (FCD) and thermo-optic (TO) effects, it is possible to map the input data of the RC to a higher dimensional space. Furthermore, by adding an external waveguide between the through and add ports of the MRR, it is possible to implement a time-delay RC (TDRC) with enhanced memory. The input from the through port is fed back into the add port of the ring with the delay applied by the external waveguide effectively adding memory. In a TDRC, the nodes (virtual) are multiplexed in time, and their respective time evolutions are detected at the drop port. The performance of MRR-based TDRC is highly dependent on the amount of nonlinearity in the MRR. The nonlinear effects, in turn, are dependent on the physical properties of the MRR as they determine the lifetime of the effects. Another factor to take into account is the stability of the MRR response, as strong time-domain discontinuities at the drop port are known to emerge from FCD nonlinearities due to self-pulsing (high nonlinear behaviour). However, quantifying the right amount of nonlinearity that RC needs for a certain task in order to achieve optimum performance is challenging. Therefore, further analysis is required to fully understand the nonlinear dynamics of this TDRC setup. Here, we quantify the nonlinear and linear memory capacity of the previously described microring-based TDRC scheme, as a function of the time constants of the generated carriers and the thermal of the TO effects. We analyze the properties of the TDRC dynamics that generate the parameter space, in terms of input signal power and frequency detuning range, over which conventional RC tasks can be satisfactorily performed by the TDRC scheme.

Neuromorphic computing hardware that requires conventional training procedures based on backpropagation is difficult to scale, because of the need for full observability of network states and for programmability of network parameters. Therefore, the search for hardware-friendly and biologically-plausible learning schemes, and suitable platforms, is pivotal for the future developments of the field. We present a novel experimental study of a photonic integrated neural network featuring rich recurrent nonlinear dynamics and both short- and long-term plasticity. Scalability in these architectures is greatly enhanced by the capability to process input and to generate output that are encoded concurrently in the temporal, spatial and wavelength domains. Moreover, we discuss a novel biologically-plausible, backpropagation-free and hardware-friendly learning procedure based on our neuromorphic hardware.

The field of artificial intelligence and machine learning (AI/ML) has experienced unprecedented growth over the last decade driven by computationally demanding applications. The computing power has been so far provided by general-purpose digital hardware such as central processing units (CPUs) and graphics processing units (GPUs). As the potential for continuous technological advancements in digital electronics is brought into question, research is focusing on alternative paradigms such as application-specific analog hardware. Both electronics and photonic analog hardware are being actively investigated with promising results showing advantages in terms of processing speed and/or energy efficiency. However, a systematic comparison of these different hardware platforms in terms of high-level computing performance is missing. In this work, we compare these hardware platforms focusing on use cases with different requirements in terms of, e.g., compute capacity, efficiency, and density. The comparison highlights current advantages and key challenges to be addressed in each field.

The specklegram analysis due to macro-bending of optical fibers has been widely employed for different sensing purposes. In this work, we mainly detect the random, multiple macro-bending loss by employing a deep learning-based convolutional neural network (CNN) namely the AlexNet model. Here, we detect the discrete losses corresponding to six macro-bends of different radii at six different locations of plastic optical fiber (POF). The proposed model can detect the macro-bending losses with 100% detection accuracy which signifies the efficacy of the proposed AlexNet model. In perspective, our results may pave the way for developing a deep-learning methodology for the smart detection of several, discrete macro-bending losses in POFs for several sensing applications.

This study focuses on using an artificial intelligence to explore metal-organic framework (MOFs) supporting the structural transformations (for instance, phase change, structural breathing, and crystal-to-crystal phase transition). Since the most MOFs possess flexible and adaptive structure, they are widely used as smart materials for optical keys, triggers, switchers, and even information encrypts. However, 100.000 potential MOFs are strongly complicated the search of specific MOF for targeted applications. Here, we report on a unique database of MOFs demonstrating the structural transformation occurring between different space groups or crystal symmetries. Using a autoencoder and classifier to predict the structural transformations, we build a link between the initial MOF structure and the potential to be switched.∗ The results pave the way to predict and design an efficient phase change MOFs for potential application in optical data processing and storage.

We propose a dual-task, all-optical processor controlled by a Machine Learning-based digital optimizer. This innovative system compensates for signal impairments in Intensity Modulation/Direct Detection (IM/DD) and mitigates bandwidth limitations, reducing the need for power-hungry Digital Signal Processing (DSP). The synergy between the optical processor and digital optimizer, operating with the Tree-structured Parzen Estimator algorithm, forms a photonic Reservoir Computing approach that enhances signal performance and reach. This versatile system can function as an all-optical equalizer and chromatic dispersion compensation filter, outperforming traditional electrical counterparts. The proposed scheme, offers a cost-effective and efficient solution for next-generation access networks.

Nowadays, sophisticated ray tracing software packages are used for the design of optical systems, including local and global optimization algorithms. Nevertheless, the design process is still time-consuming with many manual steps, taking days or even weeks until an optical design is finished. To address this shortcoming, with reinforcement learning, an agent can be trained to use ray tracing and optimization software designing an optical system. In this setting, the agent can modify the current state of the system with a predefined set of actions. One of the primary challenges is the selection of an appropriate action space. Different types of discrete and continuous action spaces are compared and their advantages and disadvantages in terms of the cumulated reward, convergence rate and resulting optical design are examined.

Photonic crystals are periodic structures with refractive index changes in one, two, or three dimensions. Due to their unique design, these crystals exhibit a photonic band gap that allows light to propagate through the structure at specific frequencies and be reflected at other frequencies. In regions where light cannot pass, known as the forbidden band gap, certain photonic states are created by deliberately creating defects in the crystal. These are called defect modes. By analyzing the dispersion curve of the defect mode, valuable information can be obtained about the behavior of light within the structure. This information includes the group velocity of the light, group velocity dispersion, and sensor sensitivity. This study proposes a two-dimensional square lattice symmetry photonic crystal design. This design arranges dielectric rods on a low refractive index material according to the square lattice symmetry. The dispersion curve of the defect mode obtained through the created line defect in the structure is investigated, and the change in group velocity of the propagating light within the structure is obtained. Increasing the sensor sensitivity is achieved by reducing the group velocity of the propagating light. Classification-based machine learning methods are employed to detect chemical substances, and the performance rates of these methods are compared for chemical substance detection.

We introduce an ML-driven optical signal processor for dispersion compensation in B5G RAN. This approach leverages a reconfigurable, energy-efficient MRR structure, effectively mitigating power fading. Our study exploits M-QAM digitally up-converted A-IFoF transmission simulation results to fiber distances up to 25km to prove the capabilities of the designed machine learning-based analog photonic processing unit. Analytical MATLAB calculations show enhanced output power, corroborated by VPI simulations demonstrating improved EVM values, including 16.9% EVM for 1GBd QPSK at 8.5GHz over 25km, meeting the 3GPP standards.

Deep neural networks (DNNs) have been successfully applied to solve complex problems, such as pattern recognition when analyzing big data. To achieve a good computational performance, these networks are often designed such that they contain a large number of trainable parameters. However, by doing so, DNNs are often very energy-intensive and time-consuming to train. In this work, we propose to use a photonic reservoir to preprocess the input data instead of directly injecting it into the DNN. A photonic reservoir consists of a network of many randomly connected nodes which do not need to be trained. It forms an additional layer to the deep neural network and can transform the input data into a state in a higher dimensional state-space. This allows us to reduce the size of the DNN, and the amount of training required for the DNN. We test this assumption using numerical simulations that show that such a photonic reservoir as preprocessor results in an improved performance, shown by a lower test error, for a deep neural network, when tested on the one-step ahead prediction task of the Santa Fe time-series. The performance of the stand-alone DNN is poor on this task, resulting in a high test error. As we also discuss in detail in [Bauwens et al, Frontiers in Physics 10, 1051941 (2022)], we conclude that photonic reservoirs are well-suited as physical preprocessors to deep neural networks for tackling time-dependent tasks due to their fast computation times and low-energy consumption.

Integrated photonic circuits offer a promising platform to implement matrix-vector multiplication in optical feedforward neural networks. The most common implementations rely on thermal phase shifters, which are inevitably susceptible to effects such as thermal and electrical crosstalk. Although deterministic, crosstalk-induced distortions have been challenging to accurately incorporate into physics-based analytical models. Additionally, analog hardware platforms suffer from fabrication deviations, that can have a significant impact on the computing performance, thus limiting scalability in implemented matrix size. In contrast, data-driven modeling techniques have shown to be promising approaches to modeling such circuits, yet they rely on black-box physics-agnostic modeling, require massive and unscalable amounts of training data, and cannot guarantee physically plausible results. Going beyond the data-driven black-box modeling techniques, but still taking advantage of the information captured by the data, we investigate the advantages of using physics-informed machine learning for photonic meshes. We analyze the ability of this approach to provide more accurate, less data-hungry, and physically plausible models for programmable photonic meshes. Moreover, we explore the potential to extract the knowledge from the trained model.

We numerically investigate a network of coupled dual-pumped degenerate optical parametric oscillators (DOPOs) with 16 spins built by Kerr microresonators for on-chip coherent Ising machines (CIMs). As a first step, the optical parametric gain in a phase-sensitive amplifier above the oscillation threshold in the silicon nitride cavity is optimized. We aim to emulate a spin glass represented by the bi-phase states with a π offset during the phase bifurcation of DOPOs. As a result, we achieve the extinction ratio of the phase-sensitive gain over 30dB with pump power as 10 dBm and detuning as 302MHz. Besides, the spatially multiplexed spins are then simulated to search for the ground state of the hexadecagon system mapping to feedback. We also study the time evolution of spins. To this end, with three different levels of initial magnetizations, we have obtained an annealing time of up to 45ns for anti-ferromagnetic and a time of up to 80ns for ferromagnetic coupling. Notably, the dynamic behavior of the integrated Ising spins for searching minimal Hamiltonian is investigated.

A simulation approach, based on the beam propagation algorithm, has been used to produce a large dataset of simulations for a MMI structure, interfaced to a MOS controlled metasurface. A machine learning approach has been used to classify the MMI configuration in terms of binary digital output for a 1x2 logical gate. This proof of concept paves the way to a more complex device class, supporting the recent advances in programmable photonic integrated circuits.

Lately, the field of optical computing resurfaced with the demonstration of a series of novel photonic neuromorphic schemes for autonomous and inline data processing promising parallel and light-speed computing. We emphasize the Photonic Extreme Learning Machine (PELM) as a versatile configuration exploring the randomness of optical media and device production to bypass the training of the hidden layer. Nevertheless, the implementation of this framework is limited to having the output layer performed digitally. In this work, we extend the general PELM implementation to an all-optical configuration by exploring the amplitude modulation from a spatial light modulator (SLM) as an output linear layer with the main challenge residing in the training of the output weights. The proposed solution explores the package pyTorch to train a digital twin using gradient descent back-propagation. The trained model is then transposed to the SLM performing the linear output layer. We showcase this methodology by solving a two-class classification problem where the total intensity reaching the camera predicts the class of the input sample.

The paper introduces an innovative object classification method for urban environments, employing distributed acoustic sensing (DAS) to address the complexities of urban landscapes. Utilizing omnipresent optical telecommunication cables, our approach involves a modified convolutional neural network (CNN) with transfer learning, achieving up to 85% accuracy. This method reuses most of the original network for feature extraction, with a final layer customized for new urban datasets – initially trained at the Brno University of Technology and then adapted to city center data. The model effectively identifies urban elements like vehicles and pedestrians, showcasing the potential of DAS for real-time classification in urban management and planning.

This communication explores an optical extreme learning architecture to unravel the impact of using a nonlinear optical media as an activation layer. Our analysis encloses the evaluation of multiple parameters, with special emphasis on the efficiency of the training process, the dimensionality of the output space, and computing performance across tasks associated with the classification in low-dimensionality input feature spaces. The results enclosed provide evidence of the importance of the nonlinear media as a building block of an optical extreme learning machine, effectively increasing the size of the output space, the accuracy, and the generalization performances. These findings may constitute a step to support future research on the field, specifically targeting the development of robust general-purpose all-optical hardware to a full-stack integration with optical sensing devices toward edge computing solutions.

Photonic reservoir computing is a neuromorphic computing framework which has been successfully used for solving various difficult and time-consuming problems. Due to its photonic nature, it offers many potential advantages such as a low-power consumption and fast processing speed. In this work, we aim to improve an already well-established design of a passive spatially distributed photonic reservoir computer, consisting of a network of waveguides connected via optical splitters and combiners. This spatially distributed architecture¹ has shown good performance on a 5-bit header recognition and an isolated spoken digit recognition task. However, this design only incorporates its nonlinearity at the photodiode in its read-out layer and is susceptible to losses within the network. Inspired by the delay-based approach to implement reservoir computing, we opt here for adding extra nonlinearity into the system to increase its nonlinear computational capacity. This is achieved by adding a single semiconductor laser as active component in an external optical delay line: light from the spatial reservoir is injected in a laser, and the optical output of the laser is then fed back to an input port of the spatial reservoir. Based on numerical simulations, we show that the nonlinear computational capacity is significantly increased by adding the feedback loop. This ultimately confirms that adding the active component can be useful for solving more complex tasks.

This paper presents a detailed investigation into the propagation loss characteristics of silicon nitride strip waveguides at an 850 nm wavelength, utilizing a random forest model. The primary aim is to optimize low-loss conditions in photonic integrated circuits (PICs). To achieve this, a systematic 2x3 full factorial design of experiments is implemented, focusing on different layers within the PIC framework. The study revolves around a critical examination of how the waveguide width influences propagation loss. Leveraging the random forest model, known for its high precision in complex data analysis, we delve into the correlation between various design elements and their impact on loss. This methodology not only aids in pinpointing the pivotal factors affecting loss but also elucidates their interplay, particularly emphasizing the role of waveguide width. One of the key contributions of this research is the identification of optimal material configurations that significantly reduce loss. This is instrumental in enhancing the efficiency of PICs, a crucial aspect for their performance in applications such as optical communications and photonic computing. Our approach uniquely combines empirical data analysis with machine learning techniques, offering a novel perspective in photonic engineering research. The findings of this study not only shed light on the complex dynamics of waveguide design but also pave the way for the development of more efficient and effective photonic systems. This research stands to make a significant impact in the field, presenting a comprehensive methodology for designing low-loss silicon nitride strip waveguides, thereby contributing to the advancement of photonic technologies.

In this study, we proposed a spatial photonic Ising machine with parallel processing. Through spatial multiplexing, the spin variables indicating multiple types of solutions were expanded in space. Individual calculations of the Ising Hamiltonian for multiple types of spin variables can be performed using the designed phase patterns. To evaluate the performance of the proposed method, the optimal solution for combinational optimization problems with 100 spin variables was determined for a certain number of iterations. Experimental results show that the frequency of the obtained optimised solution increases with the number of multiplexing steps.

In the emerging field of photonic nano-structures, the optimization of chiral meta-surfaces has emerged as a pivotal challenge, particularly for applications such as asymmetric transmission, circular dichroism (CD) spectroscopy, imaging, and spin-selective absorption. Traditional metasurface design methodologies have often been tied to laborious parameter tuning and iterative simulations, demanding both computational resources and domain expertise. This work introduces a faster approach by leveraging advanced deep-learning algorithms to streamline the optimization of chiral meta-nano surfaces. While using diatomic unit-element as the meta-surface’s building blocks, our proposed methodology harnesses the power of neural networks to predict and refine the geometrical layout of achiral nano-bars. The proposed framework is a Tandem Inverse Model (TIM) that incorporates a forward asymmetric transmission predicting neural network (ATNN) cascaded with an inverse neural network (INN). ATNN is trained in advance to enable swift and accurate prediction of the asymmetric optical behavior of meta-atoms with an MSE as low as 5.8 × 10^-4. The complete TIM assembly is then trained together while updating the weights of INN only and keeping the pre-trained ATNN part frozen. This stacked arrangement of the forward and the inverse design models successfully addresses the fundamental non-uniqueness issue suffered in the inverse design problems. With an MSE of about 3, the trained TIM model can optimize the nano-bar’s geometrical characteristics very rapidly. The suggested model, therefore, greatly accelerates the process of designing intricate chiral meta-atoms by simultaneously optimizing eight geometrical parameters in a matter of seconds. With this model, the optimal geometrical parameters of the achiral nano-bars of the meta-atom exhibited an AT of approximately 70%. Realizing such a high AT offers several uses, including lasers, optical cloaking, and electromagnetic shielding.