Data-driven tools have recently become a promising complementary approach to turbulence modeling techniques when predicting turbulent flows. The goal of these techniques is to increase the accuracy of the classical numerical strategies by restraining the computational resources employed. In this context, a combination of two data-driven tools--a sequential Data Assimilation based on the Ensemble Kalman Filter (EnKF) and the supervised Machine Learning (ML) model known as Random Forest Regression--are used to infuse physical information in a continuous Immersed Boundary Method (IBM) used to predict the turbulent channel flow for Re_{tau} = 550. The IBM initial model is a classical penalization method that accounts for the presence of the immersed body via a volume source term which is included in the Navier-Stokes equations. First, the coefficients driving the performance of the penalization method, normally selected by the user, are optimized using an EnKF data-driven strategy, where the parametric inference is governed by local and global physical information of the flow, such as the no-slip condition at the wall and the Reynolds number. Then, the ML techniques are trained using the DA results, and the black-box model obtained is integrated into the reference CFD solver. The DA-ML procedure is performed using an on-the-fly C++ library (CONES), which has been designed to couple an ensemble of numerical realizations performed using the CFD open-source code OpenFOAM with EnKF-based tools for the optimization of the IBM. The results, which are compared with a high-resolution DNS, show a significantly higher accuracy for the data-driven IBM when compared with the classical penalization tool.