Refractory complex concentrated alloys survive temperatures above 1,000°C. They're candidates for jet engine components, nuclear reactors, and hypersonic vehicles. The design space is enormous — four or more elements, each at concentrations between 5% and 35%, with oxidation behavior that depends on which oxide scale forms at the surface. Human intuition cannot navigate a four-dimensional composition space to find the alloys that form protective alumina scales.
Bejjipurapu and colleagues used Bayesian optimization with Gaussian process regression to search the space. Starting from 81 experimentally validated alloys, the active learning loop selected batches of five alloys at a time, balancing exploration (sampling unfamiliar compositions) and exploitation (refining promising regions). Each batch was synthesized, tested at 1,000°C, and the results fed back into the model.
The system converged on aluminum-containing compositions that form adherent external alpha-alumina scales with parabolic oxidation kinetics — the gold standard for high-temperature protection. Parabolic kinetics mean the oxide grows by diffusion, which slows as the scale thickens. The alloy protects itself by growing a barrier that becomes more effective over time.
What's structurally interesting is the feedback architecture. The model doesn't just predict — it decides what to synthesize next, and the experimental result retrains the model. The 81 starting points are a tiny sample of a vast space, but the loop converges because each experiment is maximally informative: chosen not for its expected performance but for its expected reduction of the model's uncertainty. The alloy that teaches the model the most is not necessarily the one that performs best.
The through-claim: in a design space too large for intuition, the question “what should I test next?” is more valuable than the question “what will work best?” The former reduces uncertainty; the latter optimizes within current knowledge. The loop that prioritizes learning over performance finds performance faster, because the bottleneck was never the answer — it was knowing where to look.