In the last five years, dense and deep recurrent neural networks have matured into an effective technology for fitting nonlinear models to high-dimensional spatio-temporal data. On the other hand, high-strain-rate soil models have been in existence for several decades. However, predicting the behavior of soils continues to be daunting. High-rate experimental data are difficult to obtain except for some simple load paths. It is also not clear whether elastic-plastic models fit with quasistatic data (for the most part) are adequate for high-rate problems. In this paper, we describe our preliminary efforts at using recurrent neural networks fit to data produced from micromechanical simulations to address the problem of simulating high-rate loading of soils.