Machine learning-based evaluation of performance of silicon nitride waveguide fabrication - Gradient-boosted forests for predicting propagation and bend excess losses

Aug 4, 2026

Upload by Liam

Study of silicon nitride strip waveguides at 850 nm focuses on propagation and bend excess loss characteristics to guide low-loss photonic integrated circuit fabrication. Gradient-boosted forests (GBF) are used to optimize manufacturing processes and accurately predict losses by modeling complex relationships between fabrication parameters and distinct loss mechanisms. A full factorial design of experiment links waveguide geometry and layer properties, including roughness and absorption, to measured optical performance. Predictive modeling supports improved PIC efficiency for photonic sensing applications.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Introduction

Experimental

Device fabrication

Device testing

Results

Gradient-boosted forest evaluation of propagation losses

Machine learning-based evaluation of performance of silicon nitride waveguide fabrication: Gradient-boosted forests for predicting propagation and bend excess losses. Jakob Wilhelm Hinum-Wagner1,2 ∗ , Samuel Marko Hoermann1,2 , Gandolf Feigl2 , Christoph Schmidt2 , Jochen Kraft1 , and Alexander Bergmann1 1 ams OSRAM Group, Tobelbaderstraße 30, 8141 Premstaetten, Austria 2Institute of Electrical Measurement and Sensor Systems, Graz University of Technology, Inffeldgasse 33/I, 8010 Graz, Austria Abstract. The propagation and bend excess loss characteristics of silicon nitride strip waveguides at an 850 nm wavelength were explored in this study. The aim was to optimize fabrication processes using machine learning, particularly gradient-boosted forests, to achieve low-loss photonic integrated circuits (PICs) and accurately predict the losses. The impact of waveguide geometry and layer properties on loss was examined using a full factorial design of experiment. These machine learning models’ predictive accuracy and ability to capture complex relationships between fabrication parameters and different loss mechanisms were assessed. Key parameters and interactions were identified, improving PIC efficiency for photonic sensing applications. 1 Introduction This study investigates integrated photonics, vital for high speed telecommunications, computing, and sensing, focusing on reducing optical losses in PICs primarily due to surface roughness and material absorption.[1] We examined Si3N4 waveguides under various conditions to assess processing compatibility. We used a full factorial design of experiment (DOE) to explore interactions between material parameters such as layer deposition techniques and geometrical parameters like waveguide width. By integrating traditional PIC measurements with statistical learning techniques such as gradient-boosted forests (GBF), this research enhances the understanding of how layer properties—such as roughness and absorption—affect PIC performance.[2] This study’s results highlight the effects of layer characteristics on optical losses and introduce a predictive maintenance model that combines GBF with DOE, thereby enhancing predictive accuracy for waveguidebased sensors.[3] 2 Experimental 2.1 Device fabrication The fabrication began with a 725 ± 15 µm thick silicon wafer, cleaned meticulously using wet chemistry to remove contaminants. The process continued with the deposition of a 2.4 µm-thick layer of SiO2 via high-density plasma enhanced vapour deposition (HD-PECVD) or a 2.2 µm-thick layer through wet chemical oxidation. For the *Corresponding author: [jakob.hinumwagner@ams-osram.com](jakob.hinumwagner@ams-osram.com) case of HD-PECVD SiO2 , chemical-mechanical polishing (CMP) was followed to ensure a reduced roughness, which can be compared to the thermally grown oxide. Subsequently, a 250 nm-thick Si3N4 waveguide layer was deposited using either low pressure chemical vapour deposition (LPCVD) or plasma enhanced chemical vapour deposition (PECVD), with continuous monitoring of thickness and stress via ellipsometry. A detailed visualization of the process flow can be seen in Figure 1 . Figure 1. Fabrication sequence includes; (1) substrate preparation, (2) SiO2 deposition,(3) CMP (when HD-PECVD oxide was used),(4) Si3N4 layer deposition,(5) photoresist coating,(6) UV exposure,(7) photoresist development,(8) Si3N4 reactive ion etching,(9) photoresist removal,(10) uppercladding deposition. Advanced deep UV lithography and a two-step reactive ion etching created the waveguide structures. The uppercladding was deposited as a 2.4 µm-thick sputtered SiO2 © The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 ([https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/)). or HD-PECVD SiO2. The DOE integral for this fabrication process organizes and e

FAQs

What losses are analyzed for silicon nitride waveguides in this study?

The study analyzes propagation loss and bend excess loss characteristics of silicon nitride strip waveguides at 850 nm.

How does the paper use machine learning to improve fabrication?

It applies gradient-boosted forests to predict optical losses and optimize fabrication processes by learning relationships between fabrication parameters and loss mechanisms.

What experimental approach is used to study how parameters affect losses?

A full factorial design of experiment explores interactions between material and geometrical parameters, such as layer deposition methods and waveguide width.

Aug 4, 2026

Upload by Liam