We develop an ensemble-averaged adjoint sensitivity analysis framework with nonlinear optimization. Starting from the conventional adjoint method, multiple patterns of random perturbations are imposed on the initial state of a dynamical system. Respective nonlinear forward simulations are conducted for each perturbed case, followed by tangential linearization and backward adjoint computation with a designated objective function. The obtained gradients are then ensemble-averaged to get the final sensitivity distribution, which is used for determining the impulsive input. This ensemble-averaging procedure smooths the explosively growing adjoint sensitivity curves, which exhibit multiple local extrema in backward time, thereby facilitating the identification of the global extrema. The proposed method is further extended with a nonlinear optimization procedure: an actuation determined from the ensemble-averaged adjoint sensitivity is used to modify the base flow, and a new round of sensitivity analysis is then performed around the updated base flow. This procedure is iteratively repeated until convergence of the objective function. During the optimization process, the actuations on base flow are determined by a weighted combination of the newly obtained sensitivity field and that from the previous iteration. The proposed method is validated through application to a series of numerical experiments on heavy rainfall modification. We select the Heavy Rain Event of July 2018 in Japan as the dataset. A target location for precipitation reduction is prescribed, and the optimal areas for applying impulsive actuation to the initial near-surface water vapor mixing ratio are examined to effectively reduce precipitation at the target. The results show that, by introducing appropriate perturbation, the ensemble-averaged adjoint sensitivity significantly enhances the control effectiveness compared with the non-ensemble approach, and that the improved sensitivity can be further refined through iterative nonlinear optimization. This study demonstrates the effectiveness of the proposed method and suggests the feasibility of modifying weather.