Computational Optics 2026

LED-based illumination has become increasingly attractive for spectroscopy, imaging, and sensing applications thanks to their spectral diversity and compactness. However, efficiently combining the emission from multiple LEDs into a single, well-defined output remains a major optical design challenge. The overlap of multiple LED sources often leads to significant étendue mismatch, non-uniform illumination, and reduced optical efficiency. We present an overview of recent advances in the optical design of multi-LED systems, from state-of-the-art methodologies to a novel compact illumination architecture employing a single freeform mirror to collect and combine the LED emission beams. Particular attention is given to the role of the freeform mirror to spatially and angularly merge different LED channels, alongside an evaluation of the system’s optical efficiency, compactness, spectral homogeneity, and robustness. This design performance is supported by experimental validation within a proof-of-concept demonstration, paving the way towards its implementation in spectroscopic sensing applications.

The design of optics has seen very significant advances over the last seven decades or so, a lot of which has been due to the tremendous increase in the power of computing & software. In 1675, Isaac Newton said that “If I have seen further, it is by standing on the shoulders of giants”. The same is true of the optical designers of my generation in that we have benefitted so much from the work of the “giants” of previous generations. The presentation will therefore include brief synopses of the contributions of a few of these giants. Some selected examples will be given of the advances over the computer age in the means of designing optics and also the developments in specific optical devices. The presentation will be dedicated to the memory of John Greivenkamp, a great educator in optics, who was to have been its joint author.

Our study explores the challenge of extending the field of view (FOV) in high-magnification optical systems without compromising image resolution or magnification. Despite recent advances in computational imaging and optical design, the trade-off between magnification and FOV remains a major limitation, primarily constrained by the finite size of camera sensors. Here, we demonstrate a novel approach to engineer the FOV by introducing a coded phase mask (CPM) into the imaging system, enabling multiplexed imaging of spatially separated object regions within a single camera frame. These results provide new insights into field-of-view engineering through phase encoding for wide-area imaging systems requiring high-resolution observation over extended areas.

This work explores Multi-Agent Reinforcement Learning (MARL) as a robust alternative to direct binary search (DBS) and gradient-based methods for the inverse design of hole-based silicon photonic devices. Addressing the combinatorial challenges of discrete optimization, we utilize the recently introduced Bandit Proximal Policy Optimization (BPPO) and Bandit Actor Critic (BAC) to decompose the design space into thousands of cooperative agents. The framework is evaluated through the design of an ultra-compact mode de-multiplexer for the Silicon-on-Insulator platform. We find that BPPO is able to outperform the strong DBS baseline without any additional hyperparameter tuning, while BAC cannot match its performance.

We present an optimized RCWA implementation based on the Hybrid-matrix (H-matrix) approach, extended to uniaxial and biaxial media with axes aligned to grating directions. The method supports planar and conical incidence and uses an enhanced submatrix strategy for reflection analysis, reducing complexity while maintaining accuracy. Numerical stability is ensured for lossy anisotropic materials, enabling multilayer computations that combine isotropic, uniaxial, and biaxial media. The H-matrix is derived analytically for substrates with different anisotropy types, eliminating singularities at specific incidence configurations. Instabilities at lossless interfaces are resolved by introducing minimal loss in critical tensor directions, ensuring stability and physically relevant behavior for complex stacks.

This study investigates the self-reconstruction behavior of a polarization singularity lattice field when partially obstructed. Numerical simulations demonstrate that the lattice exhibits robust self-healing properties during propagation. This phenomenon is driven by the non-zero net transverse component of the beam's Poynting vector. These findings suggest that such singular beams hold significant potential for applications in high-resolution microscopy and optical particle manipulation.

Optical neural networks perfectly combine the high parallelism, high-speed transmission, and low energy consumption characteristics with the powerful computing capabilities of neural networks, demonstrating unprecedented potential and advantages in the field of edge computing. However, when it comes to actual industrial application scenarios, the accuracy of optical neural networks still needs to be verified. In order to expand the application of optical neural networks, this work conducts a simulation study on EUV mask defect classification based on optical diffraction neural networks. By simulating the EUV lithography mask defect dataset and training an 8-layer optical diffraction neural network, high-precision classification of amplitude-type defects and phase-type defects was achieved. After 200 epochs of iterations, the classification accuracy of the model reached 100%, opening up a new technical route for the EUV mask defect detection problem.

Imaging through glass is plagued by obstructive reflections. This paper introduces a novel polarization-based method to separate reflections from the desired background transmission. Our approach overcomes key limitations of existing techniques when dealing with non-planar glass and complex lighting. We achieve this by fusing multiple image features to create a complexity map, which guides an adaptive segmentation of the image into localized patches. This allows for independent optimization within each patch using a mutual information minimization criterion, effectively handling spatial variations across curved surfaces. A final weighted fusion ensures seamless results. Experiments demonstrate superior performance in recovering clear background imagery through challenging surfaces like windshields, outperforming current methods in both visual and quantitative measures.

Singular optics continues to attract significant interest due to its potential for advancing optical communication, imaging, and metrology. Among polarization singularities, C-points—locations of circular polarization embedded in an elliptically polarized field—represent fundamental topological entities in vector beams. These beams exhibit robust structural stability under propagation, making them promising for optical system design. In this work, we investigate the diffraction behavior and topological features of C-point singularities using single-, double-, and triple-slit configurations. Theoretical modeling is performed using the Fresnel diffraction integral, where the field at the diffraction plane is computed from the superposition of left- and right-circularly polarized vortex components with topological charges. The resulting diffraction patterns reveal polarization-dependent interference structures that encode the local topological charge distribution. The simulated analyses demonstrate distinct diffraction signatures for different C-point morphologies, offering a new approach for their identification and control. These findings provide insight into the interaction between topology and diffraction in vector beams and suggest practical avenues for integrating singular beam structures into polarization-based optical system design, optical communication, and metrology applications.

This work proposes four sensor designs based on surface plasmon resonance that improve performance in the near-infrared region (1550 nm) with the highest possible Figure of Merit (FOM). The designs, featuring dielectrics made from titanium, zinc, and fluoride, in addition to a new dielectric–metal–dielectric configuration, allow for high sensitivities and ultra-high FOM values (up to 1856 RIU⁻¹). For all studied configurations, dielectric engineering was the most important variable that improved sensor performance. The robustness and thickness analysis also supports practical designs that can be used in real-world biomedical sensing applications.

A wide range of optical instruments based on laser radiation sources are available on the market. Many of them utilize optical systems to focus the radiation on an object. In this case, a beam of a specific wavelength is emitted by a laser and formed in the rear focal plane of the optical system—a specialized objective lens. This paper examines the design of high-aperture monochromatic optical systems for focusing laser radiation at fixed wavelengths characteristic of the most widely used modern laser sources. Optical calculations for specific lens objectives are presented, and the possibility of standardizing their individual design parameters is demonstrated.

Computational imaging and display principles are key enabling technologies with the potential to transform a wide range of future applications. New types of cameras could see through deep tissue, fog, or smoke. Fast and precise 3D scanners could improve medical diagnosis and therapy, and become essential for measuring dynamic scenes in robotic surgery, autonomous navigation, or additive manufacturing. Advances in 3D display and eye-tracking technologies could spark the next wave in AR/VR. Amidst these possibilities, understanding the fundamental physical and information-theoretical limits in computational imaging proves to be a powerful tool: Limits often appear as uncertainty relations, guiding us to optimize critical system parameters (e.g., speed or accuracy) by trading off less essential information for a given task. This talk will highlight the virtue of limits in computational imaging by discussing our recent research activities in industrial inspection, medical imaging, and AR/VR. Topics include new methods for high-accuracy eye tracking as well as high-speed, single-shot 3D metrology on challenging shiny surfaces for the inspection of metallic objects and surgical robot navigation. Moreover, I will introduce a set of techniques that use so-called “synthetic waves” for computational holographic imaging through scattering media -such as biological tissue- which also allow the capture of “light-in-flight” information without the need for pulsed lasers or fast detectors.

We present a co design framework built on the differentiable ray tracing software which provides accurate, differentiable simulations of complex optical systemswithin an optimization pipeline. We extend this framework to incorporate image reconstruction based on Wiener filters, allowing simultaneous optimization of the optical design and the associated signal processing algorithm. We present an example where this framework is used to jointly optimize a Petzval aspheric system used with a single Wiener deconvolution, with the goal of extending the depth of field while retaining satisfactory performance across the whole field.

We present a deep learning method for wavefront retrieval in double-pass optical systems, validated experimentally in a model eye. Unlike single-pass systems, double-pass imaging requires light to traverse the same optics twice—crucial for applications in human eye imaging. We trained a DenseNet-based model using simulated disk spread functions (DSFs) computed from the first 12 Zernike modes across physiologically relevant ranges. The model predicts Zernike coefficients from through-focus DSF volumes (−1 D to +1 D) using self-supervised learning by minimizing reconstruction error against the known physical forward model. Experimental validation employed a spatial light modulator to introduce controlled aberrations in a model eye with diffusive retina. Across 100 test cases, the method achieved mean reconstruction Zero-Mean Normalized Cross-Correlation (ZMNCC) of 0.049 and wavefront prediction RMSE of 0.067μm. This work demonstrates the first deep learning wavefront retrieval for double-pass systems, enabling future in-vivo human eye applications.

Advances in materials and device engineering, supported by numerical simulations, have accelerated performance optimization and reduced development costs of emerging PV. Many simulation frameworks only focus on individual sub-components of solar cells, limiting their ability to optimize the total energy yield (EY) of complex PV architectures. Therefore, we present a holistic, differentiable digital twin that calculates the EY from key input parameters. Our workflow integrates morphological modeling, optical simulations, electrical calculations, and device modeling. The differentiable design of the digital twin enables the systematic study of EY dependencies, and the framework identifies optimal design parameters to maximize the EY using a gradient-based approach.

Diffractive neural networks (DNNs) represent a powerful framework for laser beam shaping, capable of producing nearly arbitrary three-dimensional intensity distributions by cascading multiple phase masks. However, pixelwise optimization of such phase masks leads to nonideal beam shaping quality and induces high sensitivity to misalignments. Prior work showed that constraining layer training to a polynomial basis, specifically Zernike polynomials, as an intermediate training step leads to improved beam shaping quality while simultaneously reducing misalignment sensitivity, but it hinges on choosing an appropriate polynomial basis and suffers from limited local expressiveness. We instead propose using B-splines, inspired by their application in freeform optics, to represent arbitrary and locally defined continuous surfaces, avoiding the limitations of fixed polynomial bases while retaining their efficiency and robustness. We validate the method experimentally for both 2D and 3D beam shaping with cascaded spatial light modulators (SLMs), while also accounting for practical issues such as pixel crosstalk.

We develop an automatic field conversion method for FDTD simulations which enhances memory efficiency in time-reversible gradient computation for inverse design in nanohphotonics. Unlike the standard practice of saving field values in 32-bit or 64-bit precision, we achieve similar accuracy using smaller bit widths as well as interpolation. This is particularly beneficial for GPU-accelerated computing with reduced-precision data types. Integrated into the open-source, differentiable solver FDTDX, this method is essential for future large-scale simulations and optimizations in computational electromagnetics.

We present a hybrid method combining physics-informed neural operators (PINOs) with analytical multilayer modeling for efficient simulation of EUV nano-optical devices. Traditional rigorous electromagnetic solvers face prohibitive computational costs when modeling devices with thick multilayer structures, while the Fourier-domain plane-wave decomposition-based techniques require diffraction simulations from multiple illumination directions. Our approach decomposes the simulation into three components: PINO-based downward diffraction from the absorber, analytical plane-wave reflection from the multilayer stack, and PINO-based upward diffraction. The PINO model, trained across diverse unit cells and illumination directions, performs the required diffraction computations in milliseconds within a single inference pass, independent of the number of diffraction simulations, and reduces the simulation domain size by several factors compared with full-stack modeling. Validation against rigorous full-stack simulations shows excellent agreement. The approach is readily extensible to other spectral ranges, higher refractive index systems, and can naturally accommodate multiple round-trip back-reflections between absorber and multilayer structures without increasing the total simulation time.

Inverse design requires large, device-specific datasets. In this work, we show that a neural inverse model trained using dispersion curves on photonic crystal fibers (PCFs) can be transferred to Si3N4 strip waveguides using only a few dispersion points per device. Our approach freezes early PCF features, learns an input-scaler, and fine-tunes shared layers and heads with gradient normalization to balance width/height regression. On a small Si3N4 database, we recover geometry from just 5 wavelength-dispersion points, reaching R^2≈0.98 (width) and R^2≈0.91 (height) on test data. Querying the same target dispersion on both PCF and strip exposes platform feasibility limits (material/geometry), not merely model fit. The result is a sample-efficient, cross-platform inverse design recipe that repurposes a single trained model across multiple photonic platforms.

Metasurfaces enable powerful control of electromagnetic waves, supporting a large variety of applications relevant to the optics community. However, controlling their angular scattering response remains challenging due to the difficulty of achieving angular control with thin films and the absence of accurate modeling tools. Existing techniques cannot simultaneously handle different substrate/superstrate media and higher-order multipolar responses, which are common in dielectric optical metasurfaces. We present a rigorous modeling approach based on generalized sheet transition conditions that incorporates both effects and accurately predicts angular scattering. We demonstrate its use for designing metasurface angular filters that can replace optical 4f systems.

Modeling nanophotonic systems with complex or layered substrates is difficult due to effects like waveguide modes, surface plasmon polaritons or Bloch surface waves supported by the stratified background. Traditional differential methods, such as finite differences or finite elements, demand large domains and struggle with absorbing boundary conditions and perfectly matched layers since energy must be allowed to escape the computation window not only in the form of free-space scattered light, but also of guided modes. The surface integral equation method overcomes these challenges by incorporating the stratified background into the Green’s tensor, requiring discretization only of the scatterers’ surfaces and fulfilling the boundary condition perfectly, irrespective of the background complexity. This greatly reduces computational effort, though it results in dense matrices and complex Green’s tensor evaluations. This presentation will include some of the technical details associated with the method and present some experimentally-relevant situations of scatterers above or within stratified backgrounds.

Laser amplifiers can be modeled by the photon transport equation. A numerical method for the 3-dimensional solution of this equation is presented for the amplification of long laser pulses. The method allows an arbitrary time step selection independent of the spatial discretization. It can be used for the simulation of long chirped laser pulses in solid state lasers. The 3-dimensional simulation on finite volume grid allows to take into account a given pump configuration. Numerical simulation results results are presented.

The Radiation Spectrum Method (RSM) is a fast and accurate computational approach for modeling how light propagates in integrated photonic components. Unlike traditional beam propagation techniques, RSM uses a modal description of light instead of numerical grids, reducing computation time while improving physical insight. The method represents complex optical structures as a sequence of simple waveguide segments and efficiently calculates light coupling, reflection, and polarization effects. By combining high precision with fast Fourier transform acceleration, RSM provides a powerful tool for the design and optimization of advanced photonic devices in modern computational optics. We will give several results and in particular the last open source version usable on Mac OS or Windows available at this link on sourceforge : https://sourceforge.net/projects/rsmvisit/

Diffractive neural networks (DNNs) as physical representations of artificial neural networks have demonstrated remarkably flexible potential as optical systems for laser beam shaping but dynamic reconfiguration typically relies on active spatial light modulators (SLMs). We present a method to train DNNs so that rotating one or more passive phase masks switches the device among multiple, completely distinct beam-shaping functionalities, enabling dynamic beam control without SLMs. The method can also be realized experimentally with reflective DOEs produced with a unique direct laser writing approach.

Accurate modeling of EUV mask structures is essential for predicting imaging performance in next-generation lithography. This work presents a model with an analytical modal framework based on Bloch–Floquet theory to describe periodic EUV mask absorbers as coupled waveguide arrays. The approach captures lateral mode coupling and periodic boundary conditions while maintaining high computational efficiency and physical interpretability. Mode excitation coefficients and diffraction efficiencies are obtained analytically through overlap integrals and modal projections. The model investigates how absorber refractive index, geometry, and thickness influence mode behavior and diffraction characteristics. Comparison with RCWA highlights the model’s accuracy and helps identify its possible limitations.

Ray tracing enables accurate design, optimization, and tolerance analysis of complex optical systems but relies on accurate beam source modelling. Diode lasers, though efficient and compact, exhibit highly complex spatial, angular, spectral, and polarization behaviour due to emitter geometry and packaging-induced thermo-mechanical stress. This work presents multi-layer characterization of laser diode stacks, bars, and single emitters to generate digital shadows that capture these effects. The resulting source models significantly improve ray-tracing predictions, demonstrated on an ILT-designed pump engine for a satellite-based Doppler wind lidar using eight polarization-coupled high-power diode stacks with more than 9 kW peak power.

We develop a modeling framework based on multipolar decomposition that fully captures a metasurface electromagnetic response while using the minimum number of multipole moments, since excessive multipoles make modeling complex and computationally ineffective. One of the most important characteristics that determines the required number of multipole moments is the choice of the origin of the system of coordinates used for the multipolar decomposition. Here, we split multipolar contributions into even and odd parities and find their corresponding optimal scattering centers that minimize higher-order multipolar contributions. This allows us to model a metasurface, which would otherwise require higher-order multipoles, with only dipolar responses thus greatly simplifying its analysis.

In the present paper we investigate how optimization algorithm can be tailored to improve the lens design process. We replaced gradient-based optimisation methods by the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This stochastic algorithm is considered more robust and is well suited to avoid local optima often found in optical design. In addition, the algorithm is paired with an augmented Lagrangian method to incorporate constraints handling inside the computation framework. Performances are illustrated on a photographic zoom lens.

This work employs a generalized compact model for two-photon lithography (TPL) that balances speed and fidelity by embedding a semi-empirical treatment of oxygen inhibition (time-varying photosensitivity) and multi-species diffusion–reaction after exposure. We evaluate how optical settings, especially the optical field at the entrance pupil via phase and amplitude tuning, reshape the point spread function (PSF) and thus control voxel geometry. In addition, we investigate the impact of resist composition. The results provide a practical map from optical and resist parameters to printable voxel outcomes.

This work demonstrates a single-pixel microscope used with various detectors and light sources to create a multidimensional characteristic of samples. The system combines hyperspectral transmission, scattering, and emission characterization with fluorescence lifetime imaging. All images are created without the need to move the sample, therefore allowing for a cross-correlation of data. The single-pixel camera architecture also allows this setup to be affordable, opening potential for wider implementation.

Long wave infrared imaging system present a special challenge in optical design. There are few feasible materials with sufficient transparency in the IR range and at the same time the low signal intensity requires a large aperture optic and a wide operating spectrum. Reducing the size and weight of IR imaging systems is therefor of great interest in many applications. Meta-surfaces provide a great opportunity in this respect because they are smaller in size and weight then refractive lenses. At the same time they offer direct control over the wavefront shape making it easy to created aspheric and other hard to manufacture optical functions. Meta-surfaces are by nature a diffractive lens type and so have strong chromatic effects that complicate their application for a wide spectrum. In this contribution we therefor design a hybrid lens system for LWIR that combines meta-surfaces with refractive optics. Our designs show that by co-optimization of the meta-surface and refractive lenses a compact system can be designed that meets the system requirements and with closer reproduction of the target wavefront compared to a multi-level diffractive optical element.

Coherent Diffractive Imaging (CDI) is an interesting option for the metrology of EUV masks and wafers because of its contained costs and ease of implementation. The imaging step in CDI is performed algorithmically by iterating back and forth among the real and the Fourier space to estimate the missing phase. This usually implies the use of Fast Fourier Transforms (FFT) to model light propagation between planes. When the data is collected with a detector which is tilted with respect to the sample (or viceversa), the diffraction patterns are distorted at the detector plane. The distorted data cannot be straightforwardly processed by FFT algorithms, however it can be re–mapped (conical diffraction correction) onto a suitable grid for such a purpose. This mapping is performed by an interpolation step, which can be numerically expensive to evaluate and which depends on the knowledge of the tilt angle. Here, we present an approach for diffraction correction for CDI which is based on an efficient routine that estimates the tilt angle from a diffraction pattern alone.

ILT has been around for two decades, and its superior process-window has been demonstrated with wafer results by multiple fabs throughout the years. But its adoption in HVM has been limited to memory foundries and as a hot-spot repair tool for logic foundry. Three factors hindered the wide adoption of ILT in logic foundry: 1) the long computation time; 2) the large file sizes and long mask writing time; 3) scanners from tool vendors keep improving, making it possible for traditional OPC to meet the process window requirements. The situation is very different today: we have GPU and ML to address the runtime issue, and the new semi P49 standard for curvilinear shapes effectively address the file data volume, and the multiple-beam mask writer can write curvilinear mask in under one day, regardless of the mask complexity. Last but most importantly, the scaling and improvement in scanner is slowing down and is getting very expensive, prompting foundries to squeeze more performance from the existing tools without the expensive hardware upgrade or resort to multiple patterning. The adoption of ILT in HVM at logic foundries are both possible and necessary today. The author was involved in making the full-chip adoption of curvilinear mask represented by spline polygons at a leading foundry. It is anticipated full-chip free-form ILT will go to production running on GPU in 2026. In this talk, the author will review the history of ILT as someone who is an active participant, assess its current sttaus, and report the recent progress and make some bold predictions in the wake of GenAI.

Laser speckle impacts the pattern variability and stochasticity in high-end DUV lithography systems. These effects cannot be easily corrected during processing and represent a significant challenge in semiconductor manufacturing. Rigorous simulations of speckle are computationally expensive, which limits their applicability for more extensive simulations, such as source mask optimization. We show that speckle patterns can be approximated efficiently such that less computational resources are needed. This can speed up aerial image simulations with speckle by an order of magnitude in typical use cases.

Displacement Talbot Lithography (DTL) has emerged as a critical patterning solution for various photonic applications. While enabling an effectively unlimited depth of focus and a large exposure field, the unique image formation of DTL also imposes critical modeling demand. Traditional electromagnetic solvers are accurate but computationally heavy. In this work, we present a deep learning approach using Convolutional Neural Networks (CNNs) to predict DTL aerial images without direct Maxwell equation solving. This method accelerates the computation by more than a factor of 50 while maintaining high fidelity. This work significantly lowers the barrier for DTL modeling, facilitating its broader adoption in various nanomanufacturing scenarios.

We developed a methodology for reconstructing lithography lens aberrations directly from in-resist measurements of features under different rotations. Odd-order aberrations are extracted by analyzing asymmetries in two-trench structures, while even-order aberrations are determined through best-focus analysis of line/space patterns. By sampling features across multiple orientations and illumination settings, we enable a decomposition of the imaging system’s aberrations into individual aberration terms. A key enabler of this approach is the ability to efficiently simulate rotated 1D features. Conventional lithography simulators are optimized for features aligned with the principal axes, but suffer from performance degradation and edge artifacts (e.g., staircasing) when simulating off-axis features using 2D grids. To overcome this, we developed a technique that redefines the optical system in the simulator, allowing rotated features to be treated as if they were aligned with a principal axis. For isomorphic systems, this involves rotating definitions such as the Jones pupil and Zernike terms. For anamorphic systems this is far less trivial, as most simulation software doesn’t support anamorphic optics definitions under arbitrary angles. For that reason we created a way to redefine the anamorphic system into a pseudo-isomorphic imaging simulation problem. This simulation strategy enables rapid and accurate analysis of aberration sensitivities across a wide range of feature orientations, significantly reducing computational cost and enabling new capabilities. Such as the above-mentioned aberration reconstruction methods from in-resist data.

Advanced semiconductor manufacturing requires ever small features on the wafer. In recent years, extreme ultraviolet (EUV) lithography has become one of the critical technologies that enables this scaling. Its small wavelength (13.5nm) allows to print features down to 13nm (0.33NA) and recently down to 8nm (0.55NA). However, the trade-off between contrast and DoF also becomes critical. In this work, we present several possible SMO solutions to control the resist profile.

Fluorescence microscopy is essential for analyzing complex subcellular structures, yet multispectral imaging often suffers from geometric distortions caused by chromatic aberrations and from the inherently shallow depth of field (DoF) of high-NA systems. As a result, capturing fully in-focus volumetric information typically requires depth scanning, which is further hindered by scattering in thick biological specimens. Existing extended-depth-of-field (EDoF) solutions help mitigate these issues but remain limited in broadband, multispectral performance. In this presentation, we introduce an enhanced design framework that integrates end-to-end optimized full-color meta-optics, experimentally validated in a 4f fluorescence microscopy system. Our approach jointly learns a meta-optic placed at the Fourier plane and a physics-guided reconstruction algorithm, enabling simultaneous correction of defocus blur and chromatic aberrations. This co-designed optical–computational system delivers sharp, distortion-free extended-depth-of-field imaging across a broad spectral range, offering significant improvements for robust multispectral fluorescence microscopy.

Accurate optical alignment is critical to optical systems to ensure part performance. This work introduces a combined machine learning and optimization approach that can be applied to a generic system to a) model how alignment adjustments influence beam propagation and b) compute industrially relevant real-time corrections. The modular framework enables real-time control and works seamlessly with both standard robotic stages and custom actuator configurations. It also showcases generalisability by providing rapid and reliable alignment across automated testing environments, precision optics manufacturing, and research setups.

Transformer models offer a new approach to optical systems design by enabling broad, generative exploration of lens architectures beyond the abilities of traditional methods. Transformers can provide global optimizations as a starting point for simulated annealing, local optimization, and related techniques. Transformers capture long-range dependencies between surfaces, materials, and system-level constraints, making them well-suited for dynamic design problems such as movable lenses, tolerancing, and thermal variation. Combining supervised and reinforcement learning results in a more effective search compared to other machine learning approaches. Optimized Optics presents a Transformer-based model that searches for possible optical designs, changing surface shapes, thicknesses and glass types, just as language models choose words based on short- and long-term contexts. We apply this method to two zoom systems. These results demonstrate how Transformer models can complement human designers by accelerating concept generation and improving overall design-space search.

We introduce two fully differentiable light‑scattering toolkits that leverage open‑source automatic‑differentiation frameworks. PyMieDiff provides end‑to‑end differentiable Mie‑theory calculations, while TorchGDM implements the Green’s Dyadic Method in PyTorch for full‑field 2D/3D multi‑scattering simulations, supporting hybrid volume‑discretization and effective‑model approaches. Both packages enable rapid inverse‑design, optimization, and physics‑informed machine‑learning workflows in nano‑photonics, offering a scalable, open‑source alternative to traditional simulation pipelines.

Photonic inverse design (PID) has rapidly transitioned from a predominantly academic pursuit to a mainstream methodology for engineering advanced photonic devices. In this presentation, we highlight recent advancements in PID leveraging both GPU-accelerated finite-difference time-domain (FDTD) simulations and the angular spectrum method (ASM). We present comparative analyses of various optimization strategies, examining the trade-offs between traditional second-order techniques such as L-BFGS and more accessible first-order methods like ADAM.