Next-generation scientific applications in various fields are experiencing a rapid transition from traditional experiment-based methodologies to large-scale computation-intensive simulations featuring complex numerical modeling with a large number of tunable parameters. Such model-based simulations generate colossal amounts of data, which are then processed and analyzed against experimental or observation data for parameter calibration and model validation. The sheer volume and complexity of such data, the large model-parameter space, and the intensive computation make it practically infeasible for domain experts to manually configure and tune hyperparameters for accurate modeling in complex and distributed computing environments. This calls for an online computational steering service to enable real-time multi-user interaction and automatic parameter tuning. Towards this goal, we design and develop a generic steering framework based on Bayesian Optimization (BO) and conduct theoretical performance analysis of the steering service. We present a case study with the Weather Research and Forecast (WRF) model, which illustrates the performance superiority of the BO-based tuning over other heuristic methods and manual settings of domain experts using regret analysis.