Technical goals of this project are to analyze & learn from a set of existing Earth models towards assimilating a larger suite of multiscale models that are validated against data.

The analysis step includes:

(a) converting published global & regional models of North America to AVNI formats

(b) choosing an appropriate wavelet configuration (type, resolution, orientation)

(c) calculating wavelet coefficients through a forward wavelet transformation.

The learning step involves:

(a) characterizing the spectra of wavelet coefficients by its slope of decay with scale

(b) extracting features that are well correlated across models while identifying blindspots.

The assimilation step involves simulating a large suite of multiscale models with REM3D as the baseline large-scale structure by randomizing:

(a) spectral slopes to modulate power at shorter scales

(b) threshold correlation of extracted features to exclude blind spots.

The strength of heterogeneity can be expressed in terms of the RMS amplitudes at various wavelet spatial scales. Global analysis suggests a general trend of decreasing RMS amplitudes with spatial scale in the upper mantle across both V_P and V_S models (Figure 8). RMS variations at depth show stronger amplitudes for the large-scale scaling function, consistent with previous spherical-harmonic studies (e.g. Becker and Boschi, 2002).