The degree of carbon climate feedback by terrestrial ecosystems is intimately tied to the availability of moisture for photosynthesis, transpiration and decomposition. Extreme weather will be the "new normal" in a warming climate (e.g., IPCC 2007), and the vertical distribution of subsurface moisture and its accessibility for evapotranspiration will be a key determinant of the fate of ecosystems and their feedback on the climate system. The goal of this exploratory proposal is to develop algorithms of deep subsurface water dynamics and its interaction with deep root systems that could improve the CESM representation of subsurface hydrology and transpiration especially with extreme precipitation. A long time series of five years of high frequency (every 30 min) observations of water table at a research site in Northern California shows that the water tables, 18 meters below the surface, can respond in less than 8 hours to the first rains, suggesting very fast flow through micro-pores and fractured bedrock. Not quite as quickly as the water table rises after a heavy rain, the elevated water level recedes, contributing to down-slope flow and stream flow. We model the system with Richards' equation, which is a non-linear PDE, that describes the movement of liquids in unsaturated porous media and apply appropriate boundary conditions. The most crucial parameter of this PDE is the hydraulic conductivity K(θ), which describes the speed at which water can move in the underground. We show that the algorithm employed in CLM cannot capture the fast dynamics of the flow. We specify a saturation profile (i.e. the maximum allowed values of K(θ)), as a function of depth Ksat(z) and allow K(θ) to vary not only with the soil moisture saturation but also include a stochastic component which tries to capture the fracture flow. A range of data assimilation procedures has been developed to estimate the soil moisture in the subsurface θ(z,t) and also the stochastic hydraulic conductivity (θ). Moreover we attempt to estimate specific model parameters, with a dual assimilation approach, of the hydraulic conductivity function that will allow us to tune the model and better describe the observed data. Initial results are encouraging and the next steps include testing this new stochastic approach on data from other sites and ultimately implementing it into the CLM.