High-dimensional neural network potentials for accurate vibrational frequencies: the formic acid dimer benchmark

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Chemistry

Background:

• Machine learning potentials (MLPs) are increasingly used for large-scale atomistic simulations in chemistry and materials science.
• Accurately representing the potential-energy surface remains a challenge for modern MLPs, especially for complex systems.
• Systematic quality control of MLPs is often difficult for large systems.

Purpose of the Study:

• To validate high-dimensional neural network potentials using benchmark-quality vibrational frequencies.
• To assess the reliability of MLPs in reproducing subtle details of potential-energy surfaces.
• To investigate the formic acid dimer as a model system for MLP validation.

Main Methods:

• Utilized harmonic and anharmonic vibrational frequencies as sensitive probes for MLP validation.
• Employed state-of-the-art computational methods for frequency calculations.
• Compared results with high-level electronic structure theory (coupled cluster) and experimental data.

Main Results:

• High-quality vibrational frequencies were obtained for the formic acid dimer.
• The calculated frequencies showed excellent agreement with coupled cluster theory.
• The results demonstrated strong concordance with recently available stringent spectroscopic data.

Conclusions:

• Benchmark vibrational frequencies serve as a reliable method for validating high-dimensional neural network potentials.
• State-of-the-art MLPs can accurately reproduce spectroscopic properties for moderately sized systems.
• The formic acid dimer is a suitable model system for rigorous MLP quality control.