Reliable machine learning deployment across domains in electrochemical energy storage materials

Machine learning accelerates the discovery of electrochemical energy-storage materials but can fail under domain shift. We audit a fixed pretrained latent representation for porous supercapacitor materials, spanning porous carbons, metal-organic frameworks (MOFs), simulated MOFs, and external carbon/covalent triazine framework (CTF) datasets. Deployment risk is governed by latent geometry and local representational support rather than domain labels alone. Nearest-neighbor target-property discrepancy increases in extrapolative regions, while sparse neighborhoods show higher epistemic uncertainty. Combining latent distance with uncertainty yields interpretable safety maps that flag supported, cautionary, and unsupported deployment regimes, providing a reliability-first safeguard for machine learning (ML)-guided materials screening.