Constructing Spectroscopic-Accuracy Potential Energy Surfaces beyond CCSD(T) via Active and Transfer Learning: A Case...

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Spectroscopy

Background:

• Potential energy surfaces (PESs) are crucial for high-resolution spectroscopy and quantum dynamics.
• Machine learning (ML) models can fit PESs but typically require extensive data (10,000+ points) for spectroscopic accuracy.
• High-level quantum chemical methods are computationally expensive, limiting data generation.

Purpose of the Study:

• To develop an efficient ML framework for constructing accurate PESs with reduced data requirements.
• To improve the quality of ML-based PESs beyond standard high-level quantum chemical methods.
• To establish a transferable approach for generating accurate PESs for molecular complexes.

Main Methods:

• A two-step ML framework combining active learning (AL) and transfer learning (TL).
• Uncertainty-driven AL to build a baseline neural network PES at the Coupled Cluster with Singles and Doubles excitation Triples (CCSD(T)) level.
• Constrained TL to refine the baseline PES beyond CCSD(T) accuracy.

Main Results:

• The H2O-Ne system required only 601 (AL) and 59 (TL) data points.
• The transfer learning PES significantly improved agreement with experimental data.
• Root-mean-square error for a specific transition band decreased from 0.156 cm⁻¹ (baseline) to 0.022 cm⁻¹ (TL PES).

Conclusions:

• The proposed ML framework efficiently generates high-accuracy PESs for weakly bound molecular complexes.
• This approach significantly lowers the computational cost associated with obtaining accurate PESs.
• The method is transferable and applicable to various molecular systems requiring precise PESs.