<p>Predicting and controlling the impact of alloying element enrichment on hot ductility during steel scrap recycling is crucial for ensuring the product quality of recycled steel. This paper proposes a novel Physical Metallurgy-Guided Latent Space Framework (PM-LSF) model, which integrates physical metallurgy principles with machine learning techniques, and was developed alongside other machine learning algorithms to predict the reduction of area (RA), a key indicator of hot ductility. The PM-LSF model, by preprocessing inputs with physical metallurgy knowledge, utilizing an autoencoder for feature extraction, and employing an intermediate classifier, outperformed other models in terms of both prediction accuracy and model stability. This study utilized data from literature and newly conducted high-temperature tensile tests to develop and validate the models, analyzing the impact of various factors on RA. This research offers valuable insights for predicting hot ductility and optimizing alloying element control during steel scrap recycling.</p> Graphical Abstract <p></p>