Abstract
<jats:p>Multiple properties of materials and energy devices are linked by known physical relationships, yet machine-learning models commonly predict each property independently from descriptors, here termed Direct prediction. This study tests whether these relationships remain useful when their inputs are themselves model predictions. Across 16 target–route endpoints spanning solid-state ionic conductors, thermoelectric materials, liquid metal batteries and photovoltaic devices, reconstruction through predicted auxiliary quantities and known physical relationships, termed Physics-only, was less accurate than Direct for 13 endpoints, showing that predicted inputs can turn physically valid relationships into error-amplifying transformations. A residual model correcting the Physics-only estimate, termed Residual-only, recovered much of this systematic distortion but was not consistently superior to Direct and often exhibited similar sample-wise errors. PhysicsGate therefore retains Direct as an anchor and learns a bounded, sample-specific interpolation towards Residual-only using training-side out-of-fold reliability signals. PhysicsGate ranked first for 15 endpoints and within the top two for all 16, with positive mean paired gain for every endpoint. Its endpoint-macro mean paired gain was 0.105, equivalent to a 7.0% reduction in root-mean-square error relative to Direct. Gains reached 7–10% for thermoelectric targets but remained below 1% for most photovoltaic targets. Repeated-split analyses showed that sample adaptation outperformed training-selected fixed blending for thermoelectric, but not photovoltaic, targets. Performance differences were associated with auxiliary-error propagation, residual correctability, branch-error redundancy and learnable variation in branch utility. Physical relationships improved prediction when their reliability under predicted inputs was both recoverable and identifiable.</jats:p>