Material science literature typically only publishes successful syntheses, creating a lack of negative samples for training models