This work presents a machine-learning-based methodology for optimizing perovskite solar cells performance using a freely accessible synthetic dataset generated from numerical simulations of devices in SCAPS-1D. The dataset includes systematic variations of the device’s geometric, optical, and defect-related parameters, enabling an extensive exploration of the design space without the need for additional simulations. A Support Vector Regression (SVR) model with a radial basis function (RBF) kernel was trained as a surrogate model to predict power conversion efficiency (PCE) from the geometrical and optical variables of the device. For the optimization stage, a multi-objective framework based on Pareto-front analysis was implemented to balance the trade-off between maximizing PCE and minimizing prediction errors. As result, a cell architecture with a maximum PCE of 29.82% was identified, with predictive performance characterized by R² = 0.9124, MAE = 1.8470, and RMSE = 2.8773, supporting the use of SVR-RBF as a surrogate model on open synthetic data. To enhance physical interpretability, the SHAP framework was applied to quantify the influence of each optical parameter to the PCE, revealing the dominant factors associated with layer thicknesses, optical properties, and defect densities. The proposed approach demonstrates the potential of interpretable machine learning as a physics-informed design tool for high-performance perovskite solar cells, providing clear guidelines for device optimization in optoelectronic and photonic applications.