Combining low-fidelity trend information with empirical data to train models can improve their accuracy and reduce costs.