LFive Labs

We build AI agents that design semiconductor lasers.

The setup is borrowed from DeepMind's AlphaEvolve. A language model writes candidate designs and a set of physics models scores them. The best ones go back to the model as starting points for the next round. Most of our time goes into the physics models. Each one is calibrated against devices and measurements that other labs have published, and we keep some of those aside to test it on. None of the agent's designs have been built yet.

Our own laser

The first thing we're using it on is a laser we need ourselves. We want a 1.55 µm tile with at least 50% wall-plug efficiency and a single-mode beam that can be phase-locked to the tiles next to it. The best usable sources at that wavelength get around 17 to 24% once you count drivers and cooling. So far the designs that score best are photonic-crystal surface-emitting lasers, or PCSELs.

We need it for power beaming. The idea is to put lasers on a high, dry mountain and aim them at a mirror in low orbit, about 1,500 km up, which bounces the beam down to a receiver wherever power is expensive. The first customers would probably be satellites. After that, places that pay a lot for electricity, like diesel-run mines in the Andes, or Hawaii at around 40¢/kWh.

Narayanan Komerath's group at Georgia Tech worked out a version of this around 2011 and called it the Space Power Grid. They used 220 GHz millimeter waves. At 1.55 µm the optics come out nearly a thousand times smaller for the same beam spread, and from a high site the air is almost transparent (Fig. 1). Whether the economics pencil out comes down to end-to-end efficiency, and going from today's lasers to a 50% one roughly doubles it.

Fig. 1. Transmission straight up through the atmosphere, from our line-by-line model (HITRAN lines, molecular absorption only, 1 nm bins). Black is a 5,050 m site with 1.1 mm of precipitable water vapor, gray is sea level with 20 mm. The deep dips near 1.4 and 1.9 µm are water vapor. The dashed line marks 1.55 µm.

People

LFive Labs is Sebastian Alberdi and Juan Antonio Luera. The name comes from L5, the Lagrange point where Gerard O'Neill wanted to build space settlements and solar power satellites in the 1970s.

Contact

If you design lasers and want to see what the agent does with one of your problems, we'd like to hear from you. Same goes for anyone working on laser power converters or adaptive optics. Email [email protected].