In recent international exercises on numerical prediction of earthquake ground motion (EGM) in local surface sedimentary structures (LSSS), teams with the most advanced numerical‐modeling methods reached a very good level of agreement among different methods. The synthetics, however, were not sufficiently close to earthquake records. It was concluded that the structural model must be improved. Here, we apply adjoint tomography to 2D LSSS, aiming to find a model for which EGM characteristics will be sufficiently close to those determined from records. This is an important difference compared to traditional structural inversions. The methodology developed in the exploration, regional, and global scales cannot be directly applied, due to a relatively small amount of data, a relatively large initial waveform misfit, and low frequencies with respect to the size of the structure. We elaborated an inversion procedure specific for the local structures. We present a verification blind test that is closer to real‐data inversion than the standard synthetic inversions. A third party provided (a) seismograms numerically simulated for an undisclosed true structure, (b) source parameters, and (c) a homogeneous half‐space as the initial model. We demonstrate the quality of the inverted model up to the 4.5 Hz target frequency, using seismograms, waveform misfits, waveform goodness of fit (GOF), and mainly GOF for important EGM characteristics. The development of the 2D procedure, requiring much less computational load compared to the 3D procedure, is the first step. We assume that the procedure can be, in principle, applied to 3D structures after refinements, due to a 3D spatial distribution of sources and receivers.