🔬 applied physics

Lithography-free in situ photonic inverse design

This paper demonstrates a lithography-free approach to photonic inverse design by utilizing a thin-film lithium niobate free-form programmable device that can be in situ trained to perform diverse, high-performance optical functions—such as mode permutation, demultiplexing, and matrix-vector multiplication—thereby amortizing fabrication costs across multiple applications and overcoming the limitations of fixed, conventional inverse-designed devices.

Original authors: Yiqi Zhao, Martin M. Stein, Fan O. Wu, Ryotatsu Yanagimoto, Benjamin A. Ash, Mandar M. Sohoni, Logan G. Wright, Tatsuhiro Onodera, Peter L. McMahon

Published 2026-10-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yiqi Zhao, Martin M. Stein, Fan O. Wu, Ryotatsu Yanagimoto, Benjamin A. Ash, Mandar M. Sohoni, Logan G. Wright, Tatsuhiro Onodera, Peter L. McMahon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of tiny light circuits, engineers have long faced a frustrating paradox. To build a device that performs a specific task with light, such as sorting different colors or routing signals, they usually rely on a method called inverse design. This approach uses powerful computers to calculate the perfect shape for a chip, often resulting in structures so complex and counterintuitive that a human could never have drawn them. However, this method hits a hard wall when it comes to building the real thing. The computer simulation is never perfectly accurate; the microscopic imperfections of manufacturing and the messy reality of the physical world create a gap between the digital plan and the actual device. When the chip is finally made, it often performs worse than the computer predicted. To fix this, engineers must start over, fabricating a new chip, testing it, and trying again. This cycle is slow, expensive, and means that once a chip is made, its function is fixed forever. If you want a different function, you need a brand new chip.

A team of researchers at Cornell University and NTT Research has found a way to break this cycle. Instead of trying to perfect a digital simulation before building, they built a single, flexible chip first and then taught it how to perform different tasks directly in the lab. They created a thin slab of a special crystal called lithium niobate, which has the unique ability to change how it bends light when exposed to electricity. By covering this slab with a light-sensitive material, they could control the electricity flowing through it simply by shining a pattern of light onto the surface. This allowed them to turn the entire surface of the chip into a programmable landscape of roughly ten thousand tiny regions, each capable of altering the path of light passing through it.

The researchers then used a process they call "in situ" inverse design. Rather than simulating the device on a computer and hoping it works, they placed the physical chip in an optical setup and let it learn. They projected a specific pattern of light onto the chip to define a function, such as sorting different beams of light or multiplying numbers. The system measured how the light actually came out, compared it to the desired result, and then used a computer to calculate how to adjust the projected light pattern to get closer to the goal. This loop of measuring and adjusting happened in real time, allowing the device to find the perfect configuration for a specific task in just minutes. Because the chip is reprogrammed by light rather than etched with permanent grooves, the researchers could erase one function and write a completely new one onto the same physical piece of hardware in seconds.

The team demonstrated the power of this approach by turning their single device into a variety of different photonic tools. They programmed it to act as a switch that could rearrange four or eight different beams of light, moving them from one input to a specific output with high precision. They also taught it to separate different types of light patterns known as Hermite–Gaussian modes, a task that usually requires a dedicated, custom-built chip. In another experiment, they configured the device to perform matrix-vector multiplication, a mathematical operation essential for optical computing, using random matrices. The device performed these tasks with very low error, keeping unwanted signal leakage, known as crosstalk, to levels as low as negative twenty-seven decibels for the four-beam rearrangement.

Beyond just rearranging light, the researchers showed they could control how the device behaves across different colors, or wavelengths, of light. They programmed the chip to work identically whether the light was at 1500 nanometers or 1600 nanometers, creating a device that is robust against color changes. Conversely, they also programmed it to act as a splitter that separates light at 1310 nanometers from light at 1550 nanometers, sending each color to a different output. This dual capability proves that the same physical slab can be molded to either ignore color differences or exploit them, depending entirely on the light pattern projected onto it.

This work suggests a new path for the future of photonics, where the investment in making a chip is not wasted on a single function. Instead of a factory churning out thousands of slightly different chips for different jobs, a single device could be manufactured, characterized, and then reconfigured on demand to perform any number of tasks. While the current prototype is a laboratory device that requires external lasers and cameras to operate, the researchers outline a clear path to making these systems faster and more energy-efficient. By improving the materials and the way the light is projected, they believe these programmable slabs could eventually operate at speeds thousands of times faster than they do now, making them viable for real-world applications like data centers and optical neural networks. The key takeaway is that the rigid, one-time nature of traditional chip manufacturing is no longer a necessity; light can now be shaped and reshaped on a single surface, turning a static piece of hardware into a dynamic, multi-purpose tool.

Technical Summary: Lithography-free In Situ Photonic Inverse Design

Problem Statement
Photonic inverse design has proven capable of generating non-intuitive device geometries that surpass human design capabilities. However, the field faces three primary limitations:

Simulation–Reality Gap: Digital simulations of device behavior often fail to accurately predict experimental results due to fabrication imperfections, leading to suboptimal performance in fabricated devices.
Fabrication Iteration Costs: Correcting these gaps requires time-consuming and expensive fabrication cycles (hours to weeks), severely limiting the speed of design iteration.
Fixed Functionality: Once fabricated, inverse-designed devices are static; realizing a different function requires fabricating a new physical device.

While programmable photonics (e.g., Mach–Zehnder interferometers) allows for post-fabrication tuning, these platforms are typically constrained to human-designed topologies. Conversely, free-form programmable platforms exist but have historically struggled with update speeds or endurance issues (e.g., phase-change materials) that hinder gradient-based inverse design.

Methodology
The authors propose and demonstrate in situ photonic inverse design using a single, free-form programmable device. The core methodology involves:

Physical Platform: A thin-film lithium niobate (TFLN) slab waveguide covered by a silicon-rich nitride (SRN) photoconductive layer and an indium tin oxide (ITO) electrode. By projecting a 2D illumination pattern $I(x, z)$ onto the device, the photoconductor becomes locally conductive, redistributing an applied AC voltage (600 V, 4 Hz) to create a strong electric field across the TFLN. This induces a refractive index change ( $\Delta n$ ) via the electro-optic effect, with a tuning range of up to $1.7 \times 10^{-3}$ across approximately $10^4$ effective pixels. The index pattern can be reset or reconfigured in approximately 1.5 seconds.
Hybrid Optimization Loop: The authors employ a physics-aware training (PAT) algorithm that alternates between physical measurements and digital backpropagation:
1. Physical Forward Pass: The device is illuminated with a specific pattern, and input fields are coupled in. Output intensities are measured by a camera.
2. Digital Backward Pass: A differentiable digital model (based on the 2D beam-propagation method) calculates gradients. This model includes an empirical physics-based component for the index change and a convolutional neural network (CNN) to correct for illumination non-uniformities and imaging imperfections.
3. Parameter Update: The illumination pattern is updated via gradient descent to minimize a loss function defined by the desired photonic function (e.g., mode permutation, matrix multiplication).
In Situ Advantage: Because the optimization loop incorporates real-time experimental data, the process inherently compensates for the simulation–reality gap, fabrication defects, and calibration errors without requiring new physical fabrication.

Key Contributions and Results
The authors successfully demonstrated the inverse design of a single fabricated device to perform multiple distinct photonic functions, switching between them by changing the projected illumination pattern. The key results include:

Spatial Mode Manipulation (Single Wavelength, 1550 nm):
- 4 × 4 Gaussian-mode permutation: Achieved a maximum crosstalk of −27 dB.
- 4 × 4 Hermite–Gaussian mode demultiplexing: Achieved a maximum crosstalk of −13 dB.
- 8 × 8 Gaussian-mode permutation: Achieved a maximum crosstalk of −13 dB.
- 4 × 4 Matrix–Vector Multiplication (MVM): Demonstrated linear mapping of intensity-encoded vectors using random matrices.
Spectral Engineering:
- Broadband Operation: The device was optimized to perform the same 4 × 4 and 8 × 8 mode transformations across the 1500 nm to 1600 nm window with wavelength-independent functionality.
- Wavelength Demultiplexing: The device was inverse-designed to separate signals at 1310 nm and 1550 nm into different spatial outputs, achieving a crosstalk of −8.4 dB.
Efficiency: The inverse design process for each task converged in approximately 100 iterations, taking 10 to 20 minutes. Once a pattern is found, the device retains the function reliably for weeks.

Significance and Claims
The paper claims that this work establishes a new paradigm for integrated photonics where:

Simulation–Reality Gaps are Mitigated: By using experimental feedback in the optimization loop, the system dynamically corrects for physical imperfections that would otherwise degrade conventional inverse-designed devices.
Fabrication Costs are Amortized: A single fabricated device can realize and switch between many different analog functions, eliminating the need to fabricate separate devices for each function.
Simultaneous Control: The platform allows for the simultaneous manipulation of spatial and spectral degrees of freedom.
Scalability: The approach offers a pathway to scale inverse design to larger, more complex devices where the simulation–reality gap is traditionally too large for conventional digital-only design methods.

The authors note that while the current prototype is a laboratory device with limitations in index contrast, spatial resolution, and speed (pattern switching ~1.5 s), the demonstrated capability to reconfigure a single chip for diverse, high-performance functions represents a significant step toward practical, lithography-free photonic inverse design.

Technical Summary: Lithography-free In Situ Photonic Inverse Design

More like this