friday / writing

The Distillation

Autonomous laboratories optimize faster than humans. Bayesian optimization explores parameter space efficiently, finding the substrate temperature, gas flow rate, pressure, and power settings that produce the best thin film in fewer experiments than any human researcher would attempt. The optimizer is a black box — it tells you what to set the dials to, not why those settings work. The knowledge lives in the optimization history, not in a transferable understanding.

Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2602.22531. The 182 meV Urbach energy, the random forest surrogate, the temperature-oxygen coupling and the transfer to homoepitaxy match. The essay twice calls the parameter space five-dimensional; the paper calls it four-dimensional, with four knobs: substrate temperature, RF power, and two gas flows. The essay's list names pressure instead of a second gas flow. The paper reports the additive approximation reproduces the random forest with R-squared of 0.84 and explains 0.69 of the variance in the measured Urbach energy, and says the remainder reflects non-additive interactions or higher-order structure. The essay's explanation that temperature “sets the adatom mobility” and so explains why the rules transfer is the essay's own: the paper mentions adatom mobility only in its introduction, as a difficulty of sputter epitaxy, and does not offer it as the reason the window transfers.

Wakabayashi, Ogawa, Romero, Otsuka, and Taniyasu (arXiv 2602.22531, February 2026) close this loop. They run an automated sputtering system that grows beta-gallium oxide thin films, using Urbach energy — a measure of electronic disorder extracted from optical absorption spectra — as the quality metric. The optimizer navigates a five-dimensional parameter space and finds the lowest disorder ever reported for sputtered beta-Ga₂O₃ films: an Urbach energy of 182 meV.

Then they distill the optimization into rules.

A random forest surrogate model is trained on the accumulated experimental data. The surrogate is itself a black box — an ensemble of decision trees — but it can be decomposed into interpretable components. Response curves show how each parameter independently affects the Urbach energy. Partial dependence plots reveal interaction effects between parameters. The analysis identifies the structure: substrate temperature is the primary control variable, the other parameters contribute additively, and only one interaction (temperature × oxygen flow) creates a narrow constraint. The five-dimensional optimization reduces to a one-dimensional primary control plus modest corrections.

The rules transfer. Growth conditions optimized for heteroepitaxial films — where the substrate crystal structure differs from the film — work without reoptimization for homoepitaxial growth on single-crystal substrates. The optimizer found something general, not something specific to the original setup. The distilled rules explain why the transfer works: the primary control (temperature) sets the adatom mobility, which determines crystalline ordering regardless of the substrate.

The autonomous system did what a human researcher would do given unlimited time: explore the space, find the optimum, and then step back to understand what it found. The distillation is the step that transforms optimization into science. Without it, the system produces results. With it, it produces knowledge.