Alchemical harmonic approximation based potential for iso-electronic diatomics: Foundational baseline for Δ-machine...

Area of Science:

• Computational Chemistry
• Theoretical Chemistry
• Quantum Chemistry

Background:

• Predicting absolute electronic energies of molecules is crucial for understanding chemical properties.
• Existing models often struggle with accuracy and predictive power across series of related molecules.

Purpose of the Study:

• Introduce the alchemical harmonic approximation (AHA) for accurate absolute electronic energy prediction.
• Develop a model applicable to charge-neutral, isoelectronic diatomic molecules.
• Enhance machine learning efficiency for predicting molecular energies.

Main Methods:

• Combined AHA with an ansatz for electronic binding potential E(d).
• Calibrated the model using a single data point (nuclear charges Z1, Z2, and distance d0).
• Validated against reference data (pbe0/cc-pVDZ) for diatomics with 8, 10, 12, and 14 electrons.

Main Results:

• AHA demonstrates comparable accuracy to legacy potentials for single diatomics.
• AHA shows significantly better predictive power when extrapolating to entire isoelectronic series.
• Using AHA as a baseline for delta-learning reduces data requirements by an order of magnitude.

Conclusions:

• The alchemical harmonic approximation provides a robust and efficient method for electronic energy prediction.
• AHA significantly outperforms traditional models in extrapolating across isoelectronic series.
• AHA serves as an effective baseline for machine learning, drastically reducing data needs for achieving chemical accuracy.