Atomic force microscopy (AFM) is routinely used as a metrological tool among diverse scientific and engineering disciplines. A typical AFM, however, is intrinsically limited by low throughput and is inoperable under extreme conditions. Thus, this work attempts to provide an alternative with a conventional optical microscope (OM) by training a deep learning model to predict surface topography from surface OM images. The feasibility of our novel methodology is shown with germanium-on-nothing (GON) samples, which are self-assembled structures that undergo surface and sub-surface morphological transformations upon high-temperature annealing. Their transformed surface topographies are predicted based on OM-AFM correlation of 3 different surfaces, bearing an error of about 15% with 1.72× resolution upscale from OM to AFM. The OM-based approach brings about significant improvement in topography measurement throughput (equivalent to OM acquisition rate, up to 200 frames per second) and area (∼1 mm²). Furthermore, this method is operable even under extreme environments when an _in-situ_ measurement is impossible. Based on such competence, we also demonstrate the model’s simultaneous application in further specimen analysis, namely surface morphological classification and simulation of dynamic surfaces’ transformation.