News

Every photonic chip starts as an idea about light: split it evenly, route it by wavelength, twist it from one mode into another. Turning that idea into a real, etchable pattern of silicon and glass is where things get hard — not because we can’t imagine a shape that does the job optically, but because most of the shapes an algorithm imagines cannot actually be fabricated. Our new paper, “Intrinsically Design-Rule-Compliant Nanophotonic Inverse Design via Learned Generative Manifolds,” published in Laser & Photonics Reviews, tackles this mismatch head-on. Rather than teaching an optimizer to avoid fabrication violations, we taught it to never be able to draw one in the first place.

PAL members are presenting at the European Conference on Integrated Optics (ECIO) 2026, held at ETH Zurich Zentrum in Zurich, Switzerland! We are excited to share our latest work in integrated photonic computing with the ECIO community, with one oral presentation and one poster covering Integrated Photonic Computing for Electroencephalography Signal Forecasting (paper T2C.2) Spectral Photonic Encoder for Computational Tasks via Optoelectronic Recurrence (paper P12.15)

PAL members are heading to Strasbourg for SPIE Photonics Europe 2026! We are excited to share our latest work in integrated photonics and optical computing with the European photonics community. This year, we have two presentations in the Machine Learning in Photonics II track: Learned Manifold Reparameterization: A learned manifold reparameterization enabling DRC-compatible inverse design in silicon photonics Reconfigurable Deep Photonic Recurrent Networks: Reconfigurable deep photonic recurrent networks for optoelectronic information processing We look forward to seeing you in Strasbourg!

PAL members are heading to San Francisco for SPIE Photonics West 2026! We are excited to share our latest breakthroughs in integrated photonics and optical computing. This year, we have three oral presentations and two posters covering Design Rule-Compliant Inverse Design: Using deep generative networks for fabrication-compatible devices AI-Driven Optimization: Transformer-based surrogate simulator models complex photonic architectures Thermally Reconfigurable Devices: Novel reconfigurable splitters and adaptive photonic circuits Random Deep Photonic Processors: High-throughput optical inference with cascaded MZI arrays Photonic Imitation Learning: Deep photonic agentic networks for control tasks Our work at this conference focuses on enabling arbitrary optical capabilities through deep photonic network architectures and AI-enabled devices. We look forward to seeing you there!

Integrated photonics has always had a complicated relationship with polarization. The conventional solution has been polarization-diversity schemes: separate the two polarizations and route them through duplicate circuits—doubling both footprint and complexity. We decided to flip the script. Instead of duplicating circuits, we treat polarization as just another parameter to program directly into the geometry of our devices, just like we do for wavelength.

We’re thrilled to share that our latest work, “Experimental Demonstration of Ultra-Wideband Tapers, Splitters and Crossings with sub-0.1dB Loss Through Computationally Efficient and Data-Driven Eigenmode Expansion” has been accepted for presentation at CLEO Europe 2025! Rethinking Simulation Speed # Photonic devices like tapers, splitters, and crossings are foundational to modern optical systems—but traditionally, designing them with high accuracy and low loss comes at a heavy computational cost. This work presents an alternative: a data-driven eigenmode expansion (EME) framework that reduces simulation times from hours to milliseconds—without sacrificing physical accuracy. By leveraging precomputed eigenmode scattering matrices and parallel GPU processing, we model broadband light propagation in arbitrary geometries with remarkable speed.

What if you could simulate a photonic device in milliseconds instead of minutes or hours? This is no longer just an idea. With our recently published design framework ROMEO (Rapid Optical Eigenmode Optimizer), we’re beginning to see what’s possible when simulation times are drastically reduced. As photonics continues to expand into new applications, there’s growing demand for ultra-efficient, compact, and broadband photonic devices. Algorithmically optimized (or inverse-designed) structures offer a promising approach, but the physical simulations required often dominate the computational cost. To make meaningful progress, we need much faster (and still accurate) simulation techniques. That’s where ROMEO comes in.

We are proud to announce that Prof. Magden has been selected for the highly competitive 2247-D National Early-Stage Researchers Program Award, to support his research on integrated photonic machine learning systems. This program supports high-impact research projects that aim to fill critical gaps in their fields, offer fresh perspectives to the research community, and hold strong potential for major discoveries and technological breakthroughs. Prof. Magden’s project stands out for its innovative approach to developing photonic hardware that dramatically improves energy efficiency, scalability, and noise tolerance in machine learning systems.

As our digital world demands more speed and smarter sensors, the hardware under the hood—Photonic Integrated Circuits (PICs)—must become increasingly complex. But there’s a catch: designing these tiny chips to handle light in specific, ultra-fast ways is notoriously difficult and computationally expensive. Traditionally, engineers relied on physical intuition or slow, iterative simulations that could take hours or even days to design a single component. In a new study published in Nature Communications, PAL researchers have unveiled a workaround: a Highly-Scalable Deep Photonic Network platform that can design state-of-the-art optical components in under two minutes.