Transfer Learning to CCSD(T): Accurate Anharmonic Frequencies from Machine Learning Models.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Spectroscopy

Background:

• Calculating anharmonic vibrational frequencies for molecular motion assignment is computationally demanding at high quantum chemical theory levels.
• Existing methods struggle with accuracy and affordability for complex molecules.

Purpose of the Study:

• To present a practical and affordable method for computing coupled-cluster quality anharmonic frequencies.
• To enable efficient determination of molecular vibrational frequencies using machine learning.

Main Methods:

• Developed a novel approach (NN + VPT2) combining high-dimensional neural networks (PhysNet) with second-order vibrational perturbation theory (VPT2).
• Trained neural networks on potential energy surfaces (PESs) at various quantum chemical theory levels.
• Applied transfer learning (TL) for large molecules and high-level theory calculations.

Main Results:

• NN + VPT2 achieved anharmonic frequencies within 20 cm⁻¹ of experimental values for ~90% of modes (highest quality PES).
• Over 60% of modes were within 10 cm⁻¹.
• MP2-level calculations showed ~60% within 20 cm⁻¹, with outliers up to 150 cm⁻¹.
• Successfully assigned strongly interacting modes in formic acid and acetic acid.

Conclusions:

• The NN + VPT2 approach provides a cost-effective and accurate alternative for anharmonic frequency calculations.
• This method significantly advances the ability to assign molecular vibrational modes.
• The approach is applicable to a range of small- to medium-sized molecules.