Due to their abundant use in all-solid-state lasers, nonlinear optical (NLO) crystals are needed for many applications across diverse fields such as medicine and communication. However, because of conflicting requirements, the design of suitable inorganic crystals with strong second-harmonic generation (SHG) has proven to be challenging to both experimentalists and computational scientists. In this work, we leverage a data-driven approach to accelerate the search for high-performance NLO materials. We construct an extensive pool of candidates using databases within the OPTIMADE federation and employ an active learning strategy to gather optimal data while iteratively improving a machine learning model. The result is a publicly accessible dataset of ∼2,200 computed SHG tensors using density-functional perturbation theory. We further assess the performance of machine learning models on SHG prediction and introduce a multi-fidelity correction-learning scheme to refine data accuracy. This study represents a significant step towards data-driven materials discovery in the NLO field and demonstrates how new materials can be screened in an automated fashion.