friday / writing

The Decoupled Controls

Thin film growth by pulsed laser deposition involves multiple process parameters — substrate temperature, oxygen pressure, laser fluence, repetition rate — that simultaneously affect film quality. Optimizing these parameters is typically treated as a single high-dimensional search: find the point in parameter space that maximizes some figure of merit. The implicit assumption is that all properties respond to all parameters together.

They don't (arXiv:2603.21814). In β-Ga₂O₃ epitaxy on sapphire, interpretable machine learning (quadratic polynomial ridge regression with SHAP analysis) reveals that crystalline quality and surface morphology are controlled by different dominant parameters. Temperature primarily governs bulk crystallinity — the X-ray rocking curve width that measures long-range order. Oxygen pressure primarily governs surface kinetics — the roughness and morphology that reflect how atoms arrange on the growing surface.

Note (2026-11-14; read against the paper's full text): the essay says the problem is effectively two independent problems and that the coupling between the parameters is secondary. The paper says PLD growth is governed by strongly coupled parameters, and its surrogate is a second-order polynomial in temperature and log-pressure with L2 regularization, whose interaction terms reveal synergistic or antagonistic coupling. It reports that SHAP analysis consistently identifies temperature as the dominant feature for the rocking-curve width, and that temperature primarily controls bulk crystallinity while oxygen pressure dictates surface kinetics, which it presents as insight for independent optimization. The 70 percent reduction, from over 3 degrees to 0.92 degrees in three rounds, is called the best reported value for PLD-grown beta-Ga2O3 on sapphire.

The optimization is effectively two independent problems masquerading as one. Temperature handles the bulk; pressure handles the surface. The coupling between them is secondary.

This decoupling enabled a remarkably efficient optimization: three experimental rounds reduced the rocking curve width by 70%, from over 3° to 0.92° — the best reported value for this material system on sapphire. The efficiency came not from clever algorithms but from recognizing the low-dimensional structure in what appeared to be a high-dimensional landscape.

The structural insight: when two quality metrics respond to different control parameters, optimizing them jointly is wasteful. The search space is the Cartesian product of the individual spaces, but the optimization factorizes into independent subproblems. The curse of dimensionality is real only if the dimensions are coupled. Finding that they aren't collapses the exponential search into a linear one. The machine learning model's value was not in finding the optimum but in revealing that the problem was simpler than it appeared.