Evaluating transfer across modalities in autonomous materials systems

A power device needs a wide-gap semiconductor and a heat-treated Al–Mg–Si alloy to cool it, and both draw on sources that disagree: band gaps from several levels of density functional theory (DFT) and from measurement, alloy formation energies from DFT and interatomic potentials. Multi-fidelity methods combine such sources only through paired values on a common scale and do not state in advance whether one source’s guidance holds in another. We test whether the relative differences between candidates, which lead and by how much, are preserved across sources whose values disagree. PEGKi infers relative ratings of decision bodies (any rule that ranks candidates) from comparison outcomes within each source, and scores every donor-to-receiver route with a robustness certificate or a refusal. The evidence is 261,053 archived band-gap records from eight sources and an Al–Mg–Si pool where every query runs a DFT or molecular-dynamics calculation. Orderings alone recover the accuracy ladder of the DFT levels, and experiment spreads material classes about twice as widely as the generalised gradient approximation. Transfer follows the database of origin more than the functional, and routing learned from measured campaigns outperforms the fixed default at every query size. For the alloy, the score picks each potential’s room-temperature run as donor at heat-treatment temperatures, the best measured choice in five of six cases, and discounts the two potentials that predict Mg$_2$Si to be unstable. Relative margins therefore transfer where values do not, so a campaign can rate whose knowledge to reuse before committing compute or instrument time.