This AI Research Revolutionizes Silicon Mach–Zehnder Modulator Design Through Deep Learning and Evolutionary Algorithms

Network transmission capacity requirements have grown due to the popularity of Netflix and the IoT and the transition to a distributed computing and storage architecture. In short-range applications, where network costs are at a premium, it is especially difficult to meet such capacity demands. For example, high-speed optical interconnects (OIs) allow connectivity across widely dispersed hyper-scale data centers. The optical transmitter, where the electro-optical modulator plays a central role, is a key component of this type of system, both from a cost and performance standpoint. While widely used in long-haul and metropolitan systems, lithium niobate modulators show outstanding performance but cannot be efficiently integrated with the related electronics due to their huge footprint and high material costs.

In recent years, integrated photonics has been the focus of a lot of research in this area. Since its compatibility with complementary metal-oxide-semiconductor (CMOS) allows for not only the aforementioned monolithic integration with the electronic stage but also to take advantage of its fabrication know-how and the mature manufacturing infrastructure, silicon (Si) photonics has emerged as a high-potential platform for implementing low-cost and high-performance optical modulators.

The performance limits of silicon (Si) Mach-Zehnder modulators (MZMs), a key component of optical communication systems, are being tested in high-speed applications. However, typical optimization approaches require excessive time and resources to achieve a high-performance configuration due to the vast number of design factors and the complexity of modeling these devices. Researchers suggest a new approach to design using heuristic optimization and artificial neural networks to simplify the optimization process drastically.

A deep neural network model replaced the 3D electromagnetic simulation of a Si-based MZM. Then, this model was utilized to estimate the figure of merit within the heuristic optimizer (the differential evolution method). They apply this technique to CMOS-compatible MZMs and discover new configurations that improve upon previously established best practices in areas such as electro-optical bandwidth, insertion loss, and half-wave voltage. Since Si is a semiconductor, it can create phase shifters that rely on the plasma dispersion effect (PDE) by injecting and extracting free carriers. A fundamental constitutive block of in-phase and quadrature optical modulators (IQMs) requires eight electrical controls of the structure refractive index to make rib-waveguide-based phase shifters. Interferometers using micro-ring resonators (MRR), Michelson modulators (MM), and Mach-Zehnder modulators (MZM) can all use silicon phase shifters to manipulate interferometric patterns.

Though MRRs and Michelson interferometer modulators (MIMs) are small in footprint, use little energy, and are highly efficient in modulating signals, their bandwidth is severely constrained. In contrast, MZMs provide the optimum trade-off between modulation bandwidth, consumption, and insertion loss in high-speed systems despite their relatively large footprint and high power consumption. In addition to the benefits already mentioned, MZMs have greater thermal endurance and significantly less chirp in the modulated signal compared to MRRs and MIM.

Si photonics has emerged as a promising substrate for implementing MZMs because of its interoperability with CMOS. However, due to the lackluster electro-optic effects of Si, MZMs using this technology require extensive optimization, investigating as many design aspects as possible to attain their full potential. Researchers suggested an optimization strategy based on ANNs and DE to accomplish this.

Compared to conventional simulation on a general-purpose workstation, the inference time for the proposed model is seven orders of magnitude less. The application of multi-agent optimization, particularly DE, with high population numbers and the adjustment of optimization parameters, was made possible by the significant reduction in execution time. Novel MZM configurations outperformed those obtained by chance, utilizing the suggested combination of ANN modeling and DE optimization. This is the first time ANNs have been used to create integrated MZMs. The results obtained are intriguing and demonstrate the potential of the proposed design method; however, the work could be expanded to more complex MZM models, such as including the electrode-related parameters, or to test other heuristic optimization algorithms, such as particle swarm optimization or genetic algorithms. An examination of the optimized modulator’s system performance, including experimental findings, will be presented in subsequent studies.

Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 30k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.

If you like our work, you will love our newsletter..