Active learning in Gaussian process interpolation of potential energy surfaces

Area of Science:

• Computational chemistry
• Machine learning applications in science
• Quantum chemistry

Background:

• Intermolecular potential energy surfaces (PES) are crucial for understanding molecular interactions.
• Accurate PES require extensive data, often generated through computationally expensive methods.
• Current data generation methods like grid-based sampling and Latin Hypercubes (LHC) can be inefficient.

Purpose of the Study:

• To introduce and evaluate novel active learning schemes for generating training data for Gaussian process interpolation of PES.
• To compare the efficiency and accuracy of active learning against traditional sampling methods.
• To demonstrate the applicability of these schemes across different molecular systems.

Main Methods:

• Gaussian process interpolation was used to model potential energy surfaces.
• Three active learning schemes were developed and tested for data point selection.
• Performance was evaluated against Latin Hypercube (LHC) sampling for CO2-Ne, CO2-H2, and Ar3 systems.

Main Results:

• Two active learning schemes significantly outperformed LHC designs in terms of predictive error and data efficiency.
• In two of the tested molecular systems, active learning achieved error values an order of magnitude lower than LHC.
• The proposed methods demonstrated effectiveness in selecting points for de novo data generation, subset selection, and data augmentation.

Conclusions:

• Active learning schemes offer a superior alternative to traditional methods for generating PES data.
• These methods substantially reduce the number of data points required to achieve high accuracy.
• The developed procedures enhance the efficiency and accuracy of computational chemistry simulations.