Covering and loss-tolerant batch designs guarantee what autonomous campaigns sample in SiGeC semiconductors

Self-driving laboratories plan experiments in batches and lose runs to cracked pellets, tripped furnaces and cancelled compute jobs. Their planners give no guarantee about which combinations of processing variables a batch has sampled, so an interaction between two variables can go unnoticed. The hypothesis tested was that covering designs from software testing supply that guarantee at a run cost known in advance. A covering batch samples every pair of variable levels at least once, or twice so any single lost run leaves the guarantee intact; greedy follow-up batches touched every uncertain region with few runs; and eight evidence sources were assigned to composition and temperature regimes at least cost. In a pre-registered study of heat transport in the boron- or phosphorus-doped nanostructured SiGeC semiconductor (six coupled variables, 9000 combinations), a covering batch sampled every two-factor combination with 33 runs, the doubled batch kept that property under any single loss, and one-factor-at-a-time screening, which never varies two factors together, picked a recipe 33% above the best (22% under the SiC contrast measured later). On 16 molecular dynamics recipes the planner never saw, covering kept its guarantee but predicted no better than random batches, because only Ge content and temperature moved the response there. A covering-designed molecular dynamics campaign on strained SiGe device films found heat flow across a film grown on Si 12 to 18% above relaxed bulk, with heat flow along it unchanged. Covering is valuable while the relevance of each factor is unknown: an autonomous campaign can state, before its first run, which combinations its conclusions speak to.