Machine Learning Takes Some Guesswork Off the People Who Grow Light-Emitting Crystals

Picture a newer crystal grower handed an unfamiliar machine and someone else’s recipe. Before a semiconductor laser or light sensor can exist, its core has to be grown one atomic layer at a time inside a molecular beam epitaxy (MBE) machine. Getting a multi-step growth right has long depended on an experienced grower’s judgment and many trial runs, and a method developed on one machine may not work on the next. For that newcomer, a lot of costly, slow trial runs can stand between the recipe and a good sample.

A team at the Chinese Academy of Sciences’ Institute of Semiconductors and Zhejiang University built SemiEpi to carry part of that load. It uses machine learning to plan the growth, calibrate the machine, then watch the crystal surface as it grows and adjust conditions on the fly. The paper appeared in Science Advances on July 15, 2026, and the full text is open access.

The test job was indium arsenide quantum dots, nanoscale islands of semiconductor that emit light, designed to shine at 1240 nanometers. The paper describes that wavelength as suitable for gas sensing. The finished dots emitted at 1235 nm, close to target, and came out brighter and purer in color than a sample grown the conventional way.

Three things the software took off the grower’s plate:

• The arithmetic of the design. A physics model worked out what size the dots must be to emit at the target, and a machine learning model (CatBoost, trained on about 380 atomic force microscope measurements) translated that size into the dot density the growth should aim for.
• The untrustworthy thermometer. Temperature sensors read differently from machine to machine. SemiEpi heats a gallium arsenide wafer while a neural network watches live electron diffraction video for the moments the crystal surface rearranges itself. Those changes happen at known temperatures, so the system can map its own readings onto the true temperature before growth begins. Rather than trust the thermometer, it asks the crystal, which I find delightful.
• The judgment calls mid-growth. Two more neural networks read the diffraction video during growth. One judges whether the temperature is too high or too low for the target dot density, and the system adjusts it. The other decides when the dots are finished and closes the indium shutter. In the demonstration run, the system lowered the temperature by 2 degrees Celsius and closed the shutter on its own. The models were trained on data from 38 samples and reached validation accuracies of 99.1 to 99.9 percent.

The numbers on the finished dots are worth a look. SemiEpi’s sample had a linewidth (a measure of how pure the color is) of 29.13 meV and gave off about 1.6 times as much light as the conventional reference, which emitted at 1229 nm with a broader 34.70 meV linewidth. Microscope images showed about 5 x 10^10 dots per square centimeter, inside the target range; the conventional sample fell short at 3.7 x 10^10. According to the paper, five more growth runs all came out consistent with the design, and one additional experiment on a different MBE machine also hit the target.

Not every hand came off the controls. Someone still had to open the diffraction screen’s shutter by hand and adjust its power supply, steps the authors say could be brought under computer control. They also note a failure rate tied to how fast temperatures are changed.

The authors argue that calibrating each run against the crystal’s own physical transitions, then correcting in real time within a single growth, can make results more repeatable across reactors and less dependent on expert know-how, and that the approach should carry over to other compound semiconductors. That is one kind of dot, one wavelength and a single test on a second machine so far, with no finished sensors yet. But for the grower standing at a strange machine with a borrowed recipe, it is a glimpse of a colleague that reads the crystal alongside them.