Enhancing within-batch quality prediction by cyber-physical latent state models and incomplete first-principles