AI Design Boosts the Silicon Mach-Zehnder Modulator

Imagine a single-lane motorway entering a long tunnel. Inside, the road forks into two separate lanes. You are the traffic controller. To send a signal, you cannot simply stop the cars at the source; that would be far too slow. Instead, you manipulate the lanes.

To send a ‘zero’, you force cars in the left lane to slow down, just by a fraction of a second. When the lanes merge at the exit, the ‘slow’ car arrives out of sync with the ‘fast’ car from the right lane. They collide and vanish. Silence.

To send a ‘one’, you let them speed through unchanged. They merge perfectly at the exit, becoming a stronger beam of traffic. This is the principle of constructive and destructive interference.

The Silicon Mach-Zehnder Modulator Explained

In the world of fibre optics, we do not use cars; we use light. The device described above is a Silicon Mach-Zehnder Modulator. It splits a beam of light, delays one side using electricity, and recombines them to create ones and zeros.

However, there is a snag. To change the speed of light in one lane, we apply an electrical voltage. But the metal components delivering that voltage—specifically the inductors—act like rusty, sticky levers. They suffer from ‘parasitic effects’. Resistance and capacitance build up, creating a drag that limits how fast you can switch the signal on and off. Traditional engineering tries to fix this by manually tweaking the shape of these coils, but it is slow work. It is like trying to design a Formula 1 car by guessing the aerodynamics.

Enter the Transformer

A new study introduces a smarter architect. Instead of human trial-and-error, researchers employed a ‘Transformer-based’ neural network—the same type of AI architecture behind large language models. They fed this AI a dataset of over 10,000 inductor shapes simulated in high-fidelity software.

If the AI is the architect, a ‘genetic algorithm’ is the survival-of-the-fittest evolutionary process. The system generated random designs, tested them against the AI’s predictions, and bred the best ones together. This allowed the computer to explore design geometries that a human engineer might never consider.

The results were stark. The study measured a bandwidth jump from 26 GHz to 72 GHz—a 177% improvement. In system-level simulations, the new design demonstrated a 50% larger ‘eye opening’ (a measure of signal clarity) and a significant reduction in error rates. This suggests that the ‘data-driven’ approach could be the key to stabilising 100 Gbps optical interconnects, removing the bottleneck that currently throttles high-speed data traffic.