Response to Comment on "Pushing the frontiers of density functionals by solving the fractional electron problem"

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning in Science

Background:

• The DM21 model's ability to generalize beyond its training data is crucial for its applicability.
• Gerasimov et al. questioned the demonstration of DM21's fractional charge (FC) and fractional spin (FS) respecting capabilities in our previous work.
• Their critique focused on the overlap between our training set and benchmark datasets, and the validity of our generalization examples.

Purpose of the Study:

• To refute the claims made by Gerasimov et al. regarding the generalization capabilities of the DM21 model.
• To clarify the accuracy and relevance of our benchmark results and generalization examples.
• To reaffirm the overall quality and predictive power of the DM21 model.

Main Methods:

• Re-evaluation of the training set overlap with the bond-breaking benchmark (BBB).
• Detailed analysis of generalization examples presented in the paper.
• Direct rebuttal of specific points raised by Gerasimov et al.

Main Results:

• The asserted ~50% overlap between the training set and BBB is inaccurate.
• Generalization examples demonstrate DM21's robust performance outside the training set.
• The critique by Gerasimov et al. is based on flawed premises and irrelevant to the core findings.

Conclusions:

• The claims by Gerasimov et al. regarding DM21's lack of demonstrated FC/FS generalization are unsubstantiated.
• Our paper provides valid evidence for DM21's ability to generalize.
• The DM21 model maintains high quality and predictive accuracy for quantum chemical calculations.