Beyond MD17: the reactive xxMD dataset.

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• System-specific neural force fields (NFFs) are increasingly used in computational chemistry.
• The MD17 dataset is a common benchmark for developing NFF models, but it is limited to equilibrium geometries and does not represent chemical reactions.
• Many chemical reactions involve significant molecular deformations, such as bond breaking, which are not captured by existing datasets.

Purpose of the Study:

• To introduce a new dataset, the Extended Excited-state Molecular Dynamics (xxMD) dataset, to address the limitations of MD17 for chemical reaction modeling.
• To evaluate the performance of NFF models on this new dataset and assess their generalizability and extrapolation capabilities.

Main Methods:

• The xxMD dataset was generated using geometries sampled from direct nonadiabatic dynamics.
• Energies were computed using both multireference wavefunction theory and density functional theory.
• Neural force field models were assessed on the xxMD dataset.

Main Results:

• The xxMD dataset contains diverse geometries representative of chemical reactions.
• NFF models exhibited significantly higher predictive errors on the xxMD dataset compared to MD17 and its variants.
• The results indicate substantial challenges in developing NFF models with broad applicability and extrapolation capabilities.

Conclusions:

• The MD17 dataset is inadequate for developing NFF models that can accurately describe chemical reactions.
• The xxMD dataset provides a more realistic benchmark for NFF development, revealing limitations in current models.
• Further research is needed to create generalizable NFF models capable of handling complex chemical transformations and extrapolation.