Efficient prediction for high precision CO-N₂ potential energy surface by stacking ensemble DNN

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Machine Learning Applications

Background:

• High-dimensional potential energy surfaces (PES) are crucial for understanding van der Waals systems in quantum dynamics.
• Developing accurate PESs for these systems presents significant computational challenges.

Purpose of the Study:

• To establish an accurate, robust, and efficient machine learning model for constructing high-precision PESs for van der Waals systems.
• To utilize the CO-N2 system as a benchmark for developing and validating the machine learning approach.

Main Methods:

• Utilized benchmark potential energies for CO-N2 calculated via CCSD(T)-F12b/aug-cc-pVQZ.
• Developed a stacking ensemble deep neural network (SeDNN) model using four molecular structure descriptors.
• Trained and evaluated the model using 7966 benchmark potential energy points.

Main Results:

• The SeDNN model achieved high accuracy with MAE, RMSE, and R² values of 0.096, 0.163, and 0.9999 cm⁻¹, respectively.
• The predicted PES demonstrated spectroscopic accuracy for vibration spectra, showing excellent goodness-of-fit and prediction performance.
• The model accurately reproduced CCSD(T) potential energies and critical points, even when trained on reduced datasets (down to 20%).

Conclusions:

• The SeDNN model offers a promising and efficient alternative for constructing accurate potential energy surfaces for van der Waals systems.
• The developed model demonstrates superior performance and robustness, applicable to complex molecular interactions.
• This approach facilitates advancements in quantum dynamics simulations by providing reliable PESs.