An increasing number of applications require the coupling of effects from different physical fields. A leading role is played by the field of fluid-structure interaction, which investigates systems where flow and structural motion cannot be decoupled. Although these phenomena are widespread, their numerical simulation is only moderately used due to the tremendous computational cost and required expert knowledge. The high cost is caused by the iterative procedure within each time step and is aggravated by inherent numerical instabilities, increasing the number of the coupling iterations. In the state of the art, these instabilities are countered through the implementation of quasi-Newton methods, where the Jacobian of the coupled system is approximated based on previous subproblem evaluations. However, these techniques show important shortcomings, which this proposal intends to overcome by developing algorithms that automatically and dynamically set the subproblem solver tolerances, perform an automated classification and treatment of nonlinearities, and automatically learn the time-dependent behavior. In this way, I intend to drastically improve the computational efficiency and robustness of fluid-structure simulations, allowing for simulations that were previously very difficult or even impossible and boosting a more widespread use of numerical fluid-structure interaction simulations, eventually enabling advancement in the fields of renewable energy, health and many more.