Earth s solid interior harbors the long-term memory of unending transformations, such as the formation and evolution of heterogeneous domains in the mantle. These dynamic transformations operate on regions that differ in their spatial wavelengths, depth ranges and lateral extents. Seismic tomography uses waves generated by earthquakes to image the interior at global (nominal resolution > 1000 km), regional ( 500-1000 km) and local scales (< 500 km). While spectra for global VS variations are dominated by their large-scale components, regional and VP and VS spectra beneath North America have not been quantified. We analyze, learn and assimilate consistent features in the upper mantle to create a large multiscale suite with uncertainties (ATLAS3D). Our infrastructure continuously learns from the burgeoning datasets and the deluge of Earth models in the community. Multiscale models transcend spatial scales to characterize geodynamic processes (e.g. mantle mixing) and underlying causes of heterogeneity (e.g. sinking slabs, volcanism). Our approach for creating the ATLAS3D model suite leverages prior knowledge using physics-informed machine learning (PIML) and Bayesian inference. We reconcile features globally and beneath North America through: (1) analysis of features extracted using a multiscale wavelet method, (2) learning spatial- and scale-dependent blind spots based on existing models, (3) assimilating consistent features for a large suite of multiscale models, and (4) formulating a cascading validation scheme for rejecting scenarios based on agreement with data. Validation is against the vast amount of data products contributed by the community to the reference Earth model (REM3D) project. Derived datasets include attributes of body waves ( 1-20 s), surface waves ( 20-300 s) and normal modes ( 250-3300 s), typically employed in Full Spectrum Tomography. Validation also uses the band-limited waveforms ( 17-250 s) utilized in Adjoint Tomography using the Spectral Element Method. By coupling derived and processed waveform observations with predictions for arbitrary multiscale models, our results are useful for identifying regions of scientific interest, validating new techniques, planning future seismic deployments, and testing hypotheses about the Earth s interior.