materials-chemistry/ai produced the result/Nanophotonics 2023 · v2
Neural network predicts nanostructure optics to design a two-colour laser collimator
Researchers trained a convolutional neural network on simulated nanostructures to predict how each one bends light, then used its predictions to design and build a metasurface that collimates red and near-infrared laser beams.
spectrum · one line per step, placed by what the step does · bright lines used AI
Dual‐band optical collimator based on deep‐learning designed, fabrication‐friendly metasurfaces
Nanophotonics, 2023
doi:10.1515/nanoph-2023-0329 · record aix-00183 v2 · checked 2026-10-09
- AI was for
- Property prediction, Simulation surrogate
- Model family
- Convolutional neural network
- Checked by
- Experimental
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
A laser diode emits light that spreads out as it travels. A collimator is the optical part that straightens that spread into a near-parallel beam. Conventionally this means a curved glass lens. A metasurface does the same job with a flat sheet covered in millions of tiny pillars, each smaller than the wavelength of light. Every pillar, called a meta-atom, delays and dims the light passing through it by an amount set by its shape. Choose the shapes well and the sheet behaves like a lens. The difficulty is that predicting what one shape does to light requires solving Maxwell's equations numerically, which is slow, and a designer may need to compare tens of thousands of candidate shapes.
The problem grows harder when one sheet must work at two colours at once, because a single pillar has to produce the right delay at both wavelengths simultaneously. There is also a practical limit: shapes with features or gaps that are too fine cannot actually be etched into a wafer. The researchers set out to design a flat collimator that works at two wavelengths, 650 nanometres in the red and 780 nanometres in the near infrared, and that could be made with standard fabrication tools.
Where AI came in
The researchers generated roughly 20,000 random pillar shapes, drawn as 64 by 64 pixel images, and computed by conventional electromagnetic simulation how each one transmitted light. These pairs trained a convolutional neural network, a type of model that reads images, to output the transmitted light's amplitude and phase across the 650 to 780 nanometre range. Four fifths of the data were used for training and the rest held back for testing. The trained network then stood in for the simulator: paired with a shape generator, it produced a library of 80,000 free-form pillars with their predicted optical responses, a pool obtained in 91.7 seconds.
Everything after that was conventional. A rule-based filter kept the 5,300 shapes whose features and gaps were at least 60 nanometres across, and a scoring formula picked one pillar for each position on the 1 millimetre aperture. The device was then etched and measured on an optical bench, giving beam divergences of 0.11 and 0.15 degrees and transmission of 50.3 and 56.8 per cent at the two wavelengths. The network was also run once on shapes traced from electron-microscope images of the finished structure, to check how fabrication deviations mattered.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
A convolutional neural network was trained on approximately 20,000 FDTD-simulated meta-atom patterns to predict the amplitude and phase of light transmitted through free-form dielectric meta-atoms between 650 and 780 nm. The trained network was run over a pool of 80,000 generated shapes, from which a selector applying 60 nm minimum feature and gap thresholds retained 5300 fabrication-compatible patterns that covered the full 2 pi phase range at both wavelengths. Meta-atoms were assigned across a 1 mm aperture by maximising a two-wavelength figure of merit, and the resulting collimator was fabricated by electron-beam lithography and plasma etching. Measured beam divergence was 0.11 degrees at 650 nm and 0.15 degrees at 780 nm, with transmission efficiencies of 50.3 % and 56.8 % measured 50 mm from the metasurface.
How AI was used
A predictive neural network with six convolution and pooling layers followed by three fully connected layers took 64 x 64 pixel two-dimensional images of meta-atom geometries as input and output the real and imaginary parts of the transmission coefficient across the 650-780 nm spectrum. Training data were random quadrant-symmetric patterns generated in MATLAB from reference shapes such as rectangles, crosses and hollow squares at 6 nm resolution, with fixed thickness, period and measured a-Si refractive index, and labels computed by FDTD simulation in Lumerical; the set was split 80 % for training and 20 % for testing. The trained network, paired with a 2-D image generator, was then run in inference to build a library of free-form shapes with their predicted optical responses, replacing full-wave simulation of each candidate. A non-learned selector applied minimum feature size and gap thresholds to the library, and a figure of merit combining predicted transmission and phase error at the two design wavelengths was used to pick a meta-atom for each position across the collimator aperture. The network was also applied once more to geometry extracted from SEM images of the fabricated structures to check the effect of fabrication deviation.
The shape of the work
Structural · the record, drawn
no AI
Generate meta-atom pattern dataset
Obtaining raw data, whether by measurement, download or retrieval.
Approximately 20,000 random meta-atom patterns were generated from reference structures like rectangles, crosses, hollow squares, and so on with the numerical computing tool MATLAB.where the paper describes this · verbatim
no AI
Compute full-wave labels
Numerical or physics simulation, including where a learned surrogate replaces it.
The electromagnetic responses of the meta-atoms were calculated using the finite difference time domain method (FDTD)-based simulation tool Lumerical as labels.where the paper describes this · verbatim
AI
Train predictive neural network
Fitting model parameters, including fine-tuning an existing model. The AI stood in for simulation.
The meta-atoms were randomly divided into training and test datasets, with 80 % used for trainingwhere the paper describes this · verbatim
AI
Predict responses for free-form meta-atom pool
Running a trained model over new data to predict, classify or score. The AI stood in for simulation.
The fully trained PNN and 2-D image generator were used to construct a pool of meta-atoms of 80,000 free-form shapes.where the paper describes this · verbatim
no AI
Filter library with meta-atom selector
Reducing a candidate set by filtering or ranking, in a single pass.
Using this meta-atom selector, 5300 patterns were selected from the library.where the paper describes this · verbatim
no AI
Assign meta-atoms across aperture by figure of merit
Reducing a candidate set by filtering or ranking, in a single pass.
For each meta-atom position across the aperture, the FOM is calculated based on the local phase values at the two bandswhere the paper describes this · verbatim
no AI
Fabricate metasurface
Physical execution, by hand or by robot.
The metasurfaces were fabricated using electron-beam lithography and plasma etching.where the paper describes this · verbatim
no AI
Measure beam collimation and transmission
Testing outputs against ground truth.
The distance between the CMOS image sensor and the metasurface was varied from 0 to 600 mm using a linear translation stagewhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The meta-atom library that the fabricated dual-band collimator was built from was produced by the neural network's predictions rather than by full-wave simulation, so the demonstrated device depends on the model's outputs.
The meta-atom library used in this study was generated using a predictive neural network (PNN) based on a convolutional neural network (CNN) architecturewhere the paper describes this · verbatim
The total transmission efficiency measured at 50 mm from the metasurface were 50.3 % at 650 nmwhere the paper describes this · verbatim
The pool has been obtained within only 91.7 s.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of Predictive neural network (PNN)Which version of the model was used is not stated.
About this article
Record aix-00183, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error