OFC2025|Machine Learning Applications in Photonic Integrated Chips

Introduction

Photonic integrated chips (PICs) have emerged as transformative technologies in optical communications and optical computing. By using light rather than electricity, these systems can significantly reduce power consumption and achieve faster data transmission. Silicon photonics (SiPh) due to compatibility with existing semiconductor manufacturing processes, enables scalable and cost-effective production. However, fully realizing SiPh's potential involves addressing challenges in integration density, power consumption, and manufacturing complexity.

Recent research indicates machine learning (ML) can significantly enhance various stages of PIC development. This paper explores how ML accelerates PIC development in three key areas: design, fabrication, and optical characterization.

Machine Learning in Design Optimization

Modern fabrication technologies can now achieve feature sizes smaller than the wavelength of light, allowing photonics designers to create complex geometries to maximize integration density and functionality. While inverse design has become a powerful optimization tool, this iterative process is computationally intensive and often constrained by optical simulation bottlenecks.

To address these challenges, researchers have developed MetaStripNet, a deep neural network for efficiently modeling metamaterial-based photonic components. This model is particularly suitable for designing fully vertical grating couplers, which simplify packaging but require complex unit structures to achieve vertical emission while suppressing reflection.

Figure 1: (a) Concept of MetaStripNet showing nanostructures embedded in oxide cladding; (b) Fully vertical grating coupler for neural network validation; (c) Comparison of coupling efficiency and reflection simulation results using various modeling methods.

MetaStripNet employs an enhanced, environment-aware effective refractive index (nstrip) representing metamaterial sections. This approach considers not only traditional parameters like period segmentation and duty cycle but also the optical environment of the strips. By matching 3D optical responses to 2D simulations using two pre-trained deep neural networks, effective material refractive indices are obtained.

This method has been validated with grating couplers designed for SMF-28® fiber. Results closely matched 3D FDTD simulations, significantly outperforming traditional slab-layer models.

Although initial optical simulation data collection requires investment, this approach can generalize to various segmented metamaterial designs. Through global or adjoint optimization methods, the complex design space can be explored efficiently, facilitating the development of diverse grating couplers and antennas with different performance goals.

Enhancing Fabrication Precision

Manufacturing presents unique challenges in the PIC development lifecycle. Despite advances in semiconductor manufacturing technologies, PICs remain highly sensitive to dimensional variations introduced during fabrication. For high-refractive-index-contrast SiPh, even nanometer-scale deviations can cause significant spectral wavelength shifts and reduced efficiency.

Traditional methods mitigate these issues through labor-intensive calibration runs and iterative adjustments, which are both costly and time-consuming. Since these variations are influenced by multiple factors—such as pattern size, shape, and density—uniform offset layouts often inadequately address these defects.

To improve the fabrication process, researchers developed a suite of ML models collectively named PreFab. These models not only predict fabricated structures (Predictors) but also generate modified layouts in advance to correct anticipated deviations (Correctors).

Figure 2: (a) Conceptual diagram of the PreFab DNN model, showing workflows with prediction and correction functions from pattern generation to device design and corrected outputs; (b) and (c) Spectra of a topology-optimized duplex multiplexer, respectively manufactured using original and PreFab-corrected designs. Dashed lines represent 3D FDTD simulations, solid lines represent measured results.

The CNN model was trained by associating carefully selected layout patterns with corresponding fabrication SEM images. This method has been applied to topology-optimized wavelength multiplexers with a footprint of 3 × 5 μm² and a 20 nm channel spacing.

As illustrated, devices fabricated with original designs exhibited significant blue shifts in channel center wavelengths and degradation in passband shape and insertion loss. After applying the PreFab corrector model to adjust layouts, device performance closely recovered to original optimization targets.

Optical Characterization for Process Monitoring

Fabrication process monitoring is crucial for understanding structural and performance variations within and between wafers. Traditional metrology tools like SEM are unsuitable for comprehensive wafer-scale monitoring due to their local measurement capability and time-intensive nature. Additionally, these tools cannot be applied post-fabrication, limiting their practicality in process optimization.

Studies have shown geometric information of waveguides can be extracted from the optical responses of devices like ring resonators or Mach-Zehnder interferometers, often achieving higher resolution monitoring. However, traditional methods face limitations in identifying interference order and may be affected by intensity fluctuations.

To address these issues, researchers developed a novel regression model to extract waveguide cross-sectional information from ring resonator spectra. This approach distinctly identifies over 1,000 ring interference orders, allowing large bending radii to minimize curvature-induced errors.

Figure 3: (a) Layout and waveguide cross-section of a ring resonator, showing width W and height H; (b) Measured transmission spectrum; (c) Deviation between measured and predicted resonance wavelengths using extracted width and height.

Using the extracted waveguide width and height, resonance wavelengths within a 30 nm bandwidth can be predicted with an accuracy better than 5 pm. Corresponding dimensional resolution is sub-0.1 nm, independent of fitting parameters. This method offers resolution more than ten times higher than SEM and at least five times better than other waveguide-based methods.

Due to the compact resonator structure, distribution across wafers is straightforward, aiding in understanding how fabrication deviations impact PIC performance. This optical characterization approach facilitates building models that identify fabrication characteristics and inform compensatory design measures, supplementing PreFab models by providing localized wafer thickness information.

Conclusion

Machine learning has become central to PIC development, offering novel solutions across design, fabrication, and optical characterization stages. Models like MetaStripNet efficiently explore complex geometries with reduced computational demands, while PreFab enhances yields by predicting and correcting dimensional deviations. The novel optical characterization method utilizing ring resonator spectra achieves higher resolution, significantly enhancing process monitoring and feedback for manufacturing optimization.

These machine-learning-driven advancements, alongside other developments in the field, are propelling the creation of faster, more efficient, and scalable photonic devices for future communication and computing applications.

Reference

[1] D. X. Xu et al., "Advancing PIC Development Using Machine Learning: from Design to Fabrication to Optical Characterization," in OFC 2025, Optica Publishing Group, 2025, pp. Th1F.1.