Data Quality in the Fitting of Approximate Models: A Computational Chemistry Perspective

Empirical parametrization underpins many scientific methodologies including certain quantum-chemistry protocols [e.g., density functional theory (DFT), machine-learning (ML) models]. In some cases, the fitting requires a large amount of data, necessitating the use of data obtained using low-cost, and thus low-quality, means. Here we examine the effect of using low-quality data on the resulting method in the context of DFT methods. We use multiple G2/97 data sets of different qualities to fit the DFT-type methods. Encouragingly, this fitting can tolerate a relatively large proportion of low-quality fitting data, which may be attributed to the physical foundations of the DFT models and the use of a modest number of parameters. Further examination using "ML-quality" data shows that adding a large amount of low-quality data to a small number of high-quality ones may not offer tangible benefits. On the other hand, when the high-quality data is limited in scope, diversification by a modest amount of low-quality data improves the performance. Quantitatively, for parametrizing DFT (and perhaps also quantum-chemistry ML models), caution should be taken when more than 50% of the fitting set contains questionable data, and that the average error of the full set is more than 20 kJ mol^-1. One may also follow the recently proposed transferability principles to ensure diversity in the fitting set.