Embodiments of the present disclosure include methods, program products, and systems for creating a knowledge base for optical proximity correction (OPC). Methods according to the disclosure can include: fabricating a circuit using a proposed IC layout; identifying a plurality of features in an image of the fabricated circuit; predicting, based on the identifying and a predictive algorithm, whether the fabricated circuit includes a printed sub-resolution assist feature (SRAF) from the proposed IC layout; determining the predicting as being correct when the fabricated circuit includes the printed SRAF, or as being incorrect when the fabricated circuit does not include the printed SRAF; in response to the predicting being incorrect: adjusting the predictive algorithm, and flagging the fabricated circuit as incorrectly predicted; in response to the predicting being correct, flagging the fabricated circuit as correctly predicted; and storing the image of the fabricated circuit in a repository of training data.