Loading...
Technical Insight

Magazine Feature
This article was originally featured in the edition:
Issue 5 2026

Machine-learning-assisted device-to-system photonic design workflow

News

An integrated device-to-system simulation workflow combines electromagnetic modelling and machine learning to accelerate the inverse design and optimization of microring resonator demultiplexers for high-capacity data centre interconnect receivers.

BY Eugene Sokolov, Andrzej Polatynski, Hannes Lüder, Sohrab Sangini, Elias Giacoumidis, André Richter, VPIphotonics Inc., USA; VPIphotonics GmbH, Germany


The rapid development of hyperscale data centers over the past few years, driven by the massive growth of artificial intelligence (AI) and machine learning (ML) applications [1], has led to a search for new architectural solutions for optical interconnects, as traditional pluggable optical transceivers struggle to address the growing bandwidth and power consumption demands [2]. While co-packaged and near-packaged optics (CPO and NPO) solutions help meet industry requirements and improve scalability, challenges and opportunities for optimization remain [3], [4].

Shorter-reach intra-data-center interconnects (DCIs) commonly rely on direct-detection links based on wavelength-division multiplexing (WDM) and 4-level pulse amplitude modulation (PAM-4) [5]. High-capacity transmission is primarily supported through high-speed silicon microring modulators (MRMs) on the transmitter side [6]. On the receiver side, high-order microring resonators (MRRs) enable efficient, robust, and compact add-drop filters, making them particularly attractive as demultiplexers for WDM optical transmission systems [7]. In both cases, photonic integrated circuit (PIC) building blocks must be carefully designed and optimized for their target application.

However, MRM- and MRR-based PICs are highly sensitive to fabrication tolerances, design choices, and environmental variations, including thermal fluctuations [8], [9]. This sensitivity is further amplified by CPO architectures, where photonic components are placed close to high-power electronics, introducing significant thermal crosstalk. As a result, device behavior and system-level performance become tightly interdependent, making their co-optimization a crucial task that requires understanding how overall system performance depends on the individual PICs and underlying photonic devices [10]. For high-order MRR filters, this task is especially challenging: the relationship between physical design parameters, such as coupling coefficients, and system-level metrics, such as passband bandwidth and shape factor, is highly nonlinear, and conventional optimization relies on iterative electromagnetic-based device-level or circuit-level simulations, resulting in high computational cost and long design cycles. Moreover, system designers typically specify filters in terms of performance metrics rather than the physical parameters that device designers can control.

Having appropriate design automation tools at each level does not by itself resolve this problem, since the overall workflow can become fragmented, requiring manual file exchanges, custom scripting, and repeated tuning iterations between circuit- and system-level teams. We present a unified simulation framework that seamlessly integrates component characterization, custom process design kit development, and photonic circuit design with optical system simulation [11]. To illustrate the proposed workflow, we use the MRR-based demultiplexers [12] for 800G WDM links as an example and show how deep neural networks (DNNs) can be effectively employed for rapid MRR optimization and inverse design.

Device-to-system workflow
The core device-to-system workflow (Figure 1) consists of electromagnetic (EM) simulation of a photonic device, extraction of the EM solver response and import into a circuit-level compact simulation model, and validation of the resulting model in a target application [11].


Figure 1:Block diagram of the cross-layer photonic design framework based on the corresponding simulation tools [13] – [17].

At the device level, the 3D beam propagation method (BPM) and the eigenmode expansion (EME) method provide sufficient accuracy and reasonable simulation times for optimizing integrated photonic devices. The finite-difference time-domain (FDTD) method delivers the most accurate results and can serve as a rigorous final verification step, but its high computational and memory costs make it impractical for larger components and extensive optimization.

In contrast to device-level simulations, which may take several minutes for a single device configuration at a single wavelength even with fast BPM or EME methods, the corresponding circuit-level compact models are fast and typically require only a few seconds of wall-clock time. This advantage enables optimization routines and tolerance studies that require thousands of simulation runs and therefore accumulate significant total analysis time.

For devices well described by analytic or behavioral models, such as the waveguides and directional couplers that form an MRR, the most direct way to build a compact model is to extract device characteristics (e.g., effective index, coupling coefficient, loss) from EM solver results and pass them as parameters to a circuit-level model. These parameters are often wavelength-dependent arrays or functional expressions rather than single values, so both tools must support them natively. For passive linear devices that are harder to describe analytically, a full scattering-matrix (S-matrix) approach can be used instead. Both approaches require seamless, standardized interfacing between device- and circuit-level simulators, along with advanced post-processing of the exported data, including interpolation, extrapolation, regularization, and enforcement of passivity and/or causality.

To complete the design process of the PIC components, it is crucial to evaluate their performance in a transmission system. Since the device/circuit design is often addressed by a different engineering team than the system design, there is a gap between the physical device properties (such as geometrical and material properties) and the system-level performance metrics (3-dB bandwidth, shape factor, extinction ratio, insertion loss, channel spacing, symbol error rate (SER), etc.). To bridge this gap, we enhanced the workflow with ML-based inverse design, translating between device- and system-level specifications. The speed of the compact models makes it practical to generate training data from tens of thousands of simulations, enough to train accurate DNN models. In inverse design mode, the DNN takes system-level specifications as input and directly returns the corresponding compact-model or layout parameters, replacing brute-force or heuristic parameter sweeps and reducing both computation time and manual effort. Because the DNN operates on the compact models rather than underlying device physics, the framework remains accessible to users ranging from component designers to system engineers.

ML-based inverse design of MRR demultiplexers for 800G WDM links
As an example, we implement and optimize an MRR-based add-drop filter used for signal demultiplexing in a WDM PAM-4 optical DCI [5], [6]. Following the workflow described above, we first characterize the devices comprising the MRR in VPIdeviceDesigner [13]. To simulate the ring-based structures, we split them into coupling regions, where ring waveguides are coupled either to bus waveguides or to adjacent ring waveguides, and uncoupled straight or bent waveguide regions.

The waveguide is a 500 nm × 220 nm silicon wire embedded in silica cladding. A finite-difference (FD) mode solver computes the effective index of the fundamental TE mode as a function of wavelength (Figure 2a). This dispersion curve is automatically fitted with a Python expression in wavelength, which can be pasted directly into the effective index parameter of the waveguide model in VPIcomponentMaker Photonic Circuits [15].

Further, the coupling regions are characterized with BPM (Figure 2b–d). The cross-coupling power ratio is automatically extracted from the T41 (or T32) wavelength-dependent port-to-port transmission elements of the device’s S-matrix. Sweeping the gap from 50 to 80 nm across the full wavelength range gives the coupling power ratio as a function of both gap spacing and wavelength (Figure 2c). These values are exported as a text file and imported directly into the circuit-level model.





Figure 2: Effective index vs. wavelength for an isolated waveguide and corresponding fitting as a Python expression with a placeholder $L for wavelength and mode field profile in the insets (a - top left); ring-to-bus waveguide coupling region top view (b - bottom left); calculated coupling power ratio vs. gap spacing for 8 channel wavelengths (c - top right) and an example of the corresponding BPM simulation results when light propagates through a ring waveguide (d - bottom right).

With the waveguide and coupling region characterized, the circuit-level model of the third-order MRR can be assembled [12]. Compared with single-ring designs, third-order microring filters provide improved passband flatness and greater out-of-band signal rejection. Adapting the filter to the demultiplexer requires two steps. First, the resonances are aligned with the target channel frequencies using a standardized swept-tuning routine. Second, the coupling power ratios are optimized to achieve the desired passband shape (bandwidth, shape factor, and ripple) and therefore the system performance. For a symmetric three-ring filter, only two coupling ratios are independent: bus-to-ring and ring-to-ring.

These two coupling ratios, or the corresponding gaps, could be optimized directly to minimize the SER at the receiver. However, system designers do not specify filters in terms of coupling ratios or gaps, but in terms of passband metrics such as the 3-dB bandwidth, the shape factor (the ratio of the 20-dB to the 3-dB bandwidth), and passband ripple, which together determine the signal distortion and resulting SER. Even when only the 3-dB bandwidth and shape factor are used as target metrics, the mapping from coupling ratios to passband metrics is highly nonlinear and generally nontrivial.

To obtain this inverse mapping, we trained a DNN model implemented in VPItoolkit ML Framework [17] that takes 3-dB bandwidth and shape factor as inputs and returns coupling power ratios. We ran the training simulation approximately 50,000 times to build up a dataset relating filter performance to physical coupling parameters, and then used it to train a DNN model and optimize its hyperparameters. For validation, the coupling parameters predicted by the DNN were applied to the physical model, and the resulting 3-dB bandwidth and shape factor were compared with the target values originally given as DNN inputs. The prediction error was below 1%, confirming that the DNN reliably maps performance targets to device parameters.

The trained DNN model is then embedded directly into the full WDM system simulation using VPItransmissionMaker Optical Systems [16]: an 800G WDM PAM-4 optical link with 8 channels at 200 GHz spacing, modulated by cascaded MRMs, transmitted over 2 km of single-mode fiber (SMF), and demultiplexed via MRRs before signal detection and SER estimation (Figure 3a). Instead of setting the demultiplexer’s ring coupling power ratios manually, the MRR DNN model loader block determines them using the desired 3-dB bandwidth and shape factor as inputs, and a parameter controller module passes the DNN output values to the demultiplexer. The optimization can therefore operate directly on system-level filter specifications.

Using the built-in golden-section search algorithm, we then determined the 3-dB bandwidth and shape factor resulting in the lowest aggregate SER. The search converges to a 3-dB bandwidth of around 68 GHz and a shape factor of 2.4, yielding a predicted ring-to-bus coupling target of 0.174 and a ring-to-ring coupling of 0.002. The resulting DNN-optimized filter design is validated at the full system level using standard PAM-4 WDM performance metrics. The simulation typically completes a single full-system run in ~30 seconds. Results include an eye diagram and SER across channels for the final optimized configuration (Figure 3b,c).



Figure 3:800G WDM PAM-4 transmission system schematic with an MRR DNN model (a - top) and corresponding simulation results: eye diagram at the receiver for channel #6 (b - bottom left) and the SER performance for all 8 channels (c - bottom right) for the optimum MRR parameter settings.

Conclusion
To meet growing bandwidth and power-efficiency demands, new optical interconnect architectures and their optimization are needed. The described workflow combining device-, circuit-, and system-level simulation capabilities provides a streamlined data flow between different levels of abstraction – EM simulation, compact circuit models, and full transmission systems – allowing individual devices to be optimized from the perspective of system-level performance metrics. The workflow’s flexibility allows extension, as illustrated by using a DNN model for component inverse design.

The framework was applied to MRR-based demultiplexers in an 800G WDM PAM-4 link. Waveguide and coupling-region characteristics obtained from device-level EM simulations were transferred into circuit-level models of a third-order MRR filter, which were then evaluated and optimized within the full link simulation. A DNN, trained on these compact models, enabled inverse design of the ring coupling parameters from system-level filter specifications, with a prediction error below 1%. This example shows how the workflow carries device-level results all the way to system-level performance, and how separating model training from system-level optimization keeps the overall design process simple.



Logo
x