Observation-Aware Knowledge Transfer from Atomistic Models to Scientific Images

Pretrained universal machine-learning force fields (uMLFFs) provide structural knowledge for scientific image models, but transfer from three-dimensional atomistic representations to two-dimensional observations is not guaranteed to improve the image task. We study training-time transfer from a frozen atomistic encoder while retaining image-only inference, using a local spatial objective that associates projected atom positions with image features. In a DenseNet AtomVision comparison, invariant MLFF supervision increases native accuracy from 79.43% to 80.09% but increases the measured rotation loss from 65.49% to 72.01%. In a matched fixed-tip STM study, an SO(2) image backbone has a lower rotation loss than a CNN backbone, while invariant MLFF supervision does not provide a consistent additional reduction. These results support conditional structural knowledge transfer and show why native task quality and rotation loss must be evaluated separately.