Predicting beach morphodynamics over multiple decades for coastal hazard risk analyses requires accurate and efficient computation of beach recovery following highly energetic erosive storms. However, process-based sediment transport modeling of long duration, lower energy conditions generally requires time-intensive numerical models that often exhibit limited skill unless they are calibrated with high-fidelity, site-specific measurements that are difficult, expensive, and generally unavailable. As an alternative, data-driven approaches that leverage machine learning (ML) of physical processes can be used to create a surrogate model from observations. Although data-driven methods are also vulnerable to overfitting to closely match observed data, this work demonstrates how physical principles can be embedded into data-driven emulators to predict bed elevation change as a function of cross-shore coordinate, or Δz(x). Specifically, the methodology for building an ML model that incorporates spatial dependence between cross-shore bed elevations, temporal dependence between consecutive observations, and conservation principles is presented. Nearly a decade of beach profile elevations and offshore wave statistics collected by the U.S. Army Corps of Engineers Field Research Facility (FRF) in Duck, North Carolina are leveraged to demonstrate the proposed methodology.