We propose an efficient approach for the early forecast of long‐period (> 3–10 s) ground motions generated in sedimentary basins by large earthquakes based on the data assimilation of observed ground motions and finite‐difference method simulations of seismic wave propagation in a 3‐D heterogeneous structure. This approach uses the dense K‐NET and KiK‐net nationwide networks in Japan and a high‐performance supercomputer to perform forecasts using the assimilated wavefields at speeds much faster than the actual wave propagation speed. Therefore, an early alert can be issued prior to the occurrence of strong motions in basins due to large, distant earthquakes. We validated the effectiveness of this data‐assimilation‐based forecast approach via numerical tests for the early forecast of long‐period ground motions in central Tokyo using the observed waveform data from the Mw6.6 2007 Off Niigata and Mw9.0 2011 Off Tohoku earthquakes.

Long-period (LP) ground motions with periods of~2-10 s caused by large earthquakes are strongly amplified in sedimentary basins, posing serious threats to modern cities to cause resonance and damage to skyscrapers, oil storage tanks, long-span bridges, and other structures with long natural periods. Since the LP ground motions are composed of surface waves traveling for longer distances with much slower speeds than body waves, an alert could be issued by detecting the spread of strong ground motion near the source and before large ground motions occur in distant basins. In this study, an early forecasting of LP ground motions is proposed based on data assimilation of observed ground motions obtained by high-density, nationwide seismic networks and a computer simulation of the seismic wave propagation using a high-resolution subsurface structure model. The shaking of the LP ground motions in the distant basins is immediately forecasted by high-speed supercomputers from current data-assimilated wavefield. In addition, the forecasting speed is significantly improved by using Green's functions, representing the wave propagation response from the data assimilation stations to a forecast target site. The effectiveness of this data assimilation-based forecasting of LP ground motions in conjunction with using Green's functions is demonstrated by numerical tests using the observed strong motion records from the 2004 Off Kii Peninsula earthquake in Japan (Mw 7.4) together with synthetic seismograms of anticipated large earthquake scenarios in the Nankai Trough.