LEGO-like crystals add light control to silicon chips

Photonic chips are held back less by the idea than by the materials. Silicon photonics is a mature, well-funded platform, and it is excellent at guiding light. But silicon cannot do everything a light-based chip needs: it modulates light only weakly, it cannot easily make a non-reciprocal isolator, and it does not detect light well at the wavelengths that matter. For years the answer has been to grow better materials directly on silicon, an integration problem that fights the chemistry of the two substrates. A paper in Nature in October, by Yuan Meng, Wenbo Mao and colleagues at Washington University in St. Louis, led by Sang-Hoon Bae and Lan Yang, proposes the opposite: grow the crystals separately, then transfer them on like LEGO bricks.

The framework is heterogeneous integration of single-crystalline nanomembranes. The team frees ultrathin, freestanding films of the crystals they want and places them onto prefabricated silicon and silicon-nitride photonic circuits, wherever a specific function is needed. Because the crystals are grown on their own ideal substrates first, they avoid the defects that come from forcing them to grow on silicon. Because they are then placed where they are wanted, the designer is free to pick a material for each job rather than compromise on one.

The paper demonstrates the range of the toolbox. Barium titanate nanomembranes give strong electro-optic modulation, letting a voltage switch a light signal efficiently. Cobalt ferrite membranes add non-reciprocal light control, the on-chip equivalent of an isolator that lets light travel one way and not the other. Gallium arsenide and gallium nitride membranes, integrated side by side on silicon nitride, detect light across wavelengths from the ultraviolet to the near-infrared on the same chip, using different detectors in different places. Stacking barium titanate and cobalt ferrite on the same silicon micro-ring resonator then combines electric and magnetic control in one device.

The significance is modularity. Photonic integrated circuits are being chased hard for the next generation of computing, because moving data with photons is faster and less power-hungry than moving it with electrons, and because artificial-intelligence systems need optical links at enormous bandwidth. The bottleneck has been that every new function demanded a new materials process integrated into a fab that was never designed for it. A library of transferable membranes means the functions can be mixed and matched after the circuit is made, like components on a breadboard.

The honest limits are the usual integration challenges. Transfer printing ultrathin crystals onto patterned wafers at scale is not the same as doing it in a research lab, and the interfaces between membrane and waveguide still have to be made low-loss and reliable. This is a demonstration of a materials framework and a set of devices, not a manufacturing line. But the principle is what matters: by separating growth from integration, the team has decoupled the choice of material from the constraint of the substrate.

For a field that has spent a decade trying to make silicon do jobs it was never good at, the LEGO metaphor is liberating. The chip becomes a platform on which the right crystal can be dropped for each function, whether that function is modulating, isolating or detecting light. If the transfer process scales, the photonic chip of the future will not be made of silicon alone. It will be a stack of carefully chosen crystals, each one doing what it does best.