SmartSim: An Uncertainty-Guided Active Learning System for Reducing Expensive Simulator Evaluations

High-fidelity physical simulators are computationally expensive, which limits the number of evaluations that can be run when exploring a parameter space. We present SmartSim, an active learning system that couples a probabilistic surrogate model (Gaussian Process) with a high-fidelity simulator, in which an acquisition function decides at each iteration which point in parameter space is worth evaluating with the real simulator. We formalize the system, including a stopping mechanism based on the maximum uncertainty observed over a candidate region. On the synthetic Branin benchmark, and with a multi-seed protocol, active learning reaches a target error in 22.1 ± 2.7 real evaluations (15 seeds), while an equivalent random sampling baseline needs a median of 762 evaluations (30 seeds, budget extended to 2000 to eliminate budget censoring; 3 of 30 seeds still fail to converge)-a real reduction of approximately 97%, far above the 65% that a fixed budget of 60 evaluations would misleadingly suggest. 94.7% of the model's predictions fall within a µ ± 2σ interval. We further explicitly evaluate the uncertainty-based stopping criterion (H2): it works, but only once the threshold τ is calibrated to the real scale of the trained kernel rather than the intuitive scale of the problem's errora design finding we report together with the measured trade-off between evaluations saved and calibration preserved. Code is publicly available https://github.com/rainvare/modelo-del-mundo