Machine Learning Models of Vibrating H₂CO: Comparing Reproducing Kernels, FCHL, and PhysNet

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Atomistic simulations are crucial for understanding molecular behavior.
• Machine learning (ML) offers potential for enhancing simulation accuracy and efficiency.
• Formaldehyde serves as a benchmark system for evaluating ML models in molecular simulations.

Purpose of the Study:

• To comparatively assess state-of-the-art ML models for atomistic simulations.
• To evaluate the performance of deep neural networks (NNs), reproducing kernel Hilbert space (RKHS+F), and kernel ridge regression (KRR).
• To investigate the predictive capabilities of these models for energies, atomic forces, and vibrational frequencies.

Main Methods:

• Developed and trained ML models including PhysNet (NN), RKHS+F, and FCHL (KRR) using formaldehyde data.
• Generated learning curves for energies and atomic forces.
• Performed finite-temperature molecular dynamics (MD) simulations.
• Investigated transfer learning (TL) from B3LYP to CCSD(T)-F12 using PhysNet.

Main Results:

• ML models rapidly converge to accurate predictions for energies and forces with modest training sets.
• Predictive power for energy extrapolation increases from NNs to RKHS+F and KRR.
• PhysNet and FCHL show high accuracy for harmonic vibrational frequencies.
• MD simulations yield comparable infrared spectra, with limitations in high-frequency modes.
• ML models can identify convergence issues in reference electronic structure calculations.

Conclusions:

• ML models demonstrate high accuracy and efficiency for atomistic simulations of formaldehyde.
• The choice of ML method impacts performance for different properties and extrapolation tasks.
• Transfer learning offers a pathway to improve data efficiency in ML-based simulations.