LLNL-ToFi applies the resolution matrix from the LLNL-G3D-JPS seismic tomography model (Simmons et al., 2015) to arbitrary velocity models through a single sparse matrix-vector multiplication. It is a modern Python reimplementation of the original Fortran-based LLNL_ToFi code by Bernhard Schuberth at LMU Munich.
The LLNL-G3D-JPS model provides a resolution matrix that encodes how the seismic imaging process smears and distorts the true Earth structure. By applying this operator to a geodynamic velocity model, we can produce a filtered prediction that is directly comparable to the published tomography — accounting for the uneven data coverage and regularisation inherent in seismic inversions.
The package operates on 44 depth layers spanning from the crust to the core-mantle boundary (over one million grid points), supports both geographic and Cartesian coordinate systems, and handles automatic downloading and caching of the resolution data. It also fixes a boundary indexing bug present in the original Fortran code that affected a small number of matrix entries.
Like SRTS, LLNL-ToFi integrates naturally with G-ADOPT workflows for quantitative model-data comparison in mantle convection inversions.
Source: github.com/g-adopt/llnltofi