Multitask methods for predicting molecular properties from heterogeneous data.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Chemistry

Background:

• Generating accurate molecular property data is a significant challenge for training predictive models.
• Existing methods often require extensive, high-fidelity data, increasing computational costs.
• Surrogate models are crucial for accelerating molecular property prediction.

Purpose of the Study:

• To demonstrate that multitask Gaussian process regression can overcome data generation bottlenecks.
• To leverage both expensive (coupled-cluster) and inexpensive (density functional theory) data sources.
• To reduce the cost of generating training data for molecular property prediction models.

Main Methods:

• Employed multitask Gaussian process regression to integrate data from coupled-cluster (CC) and density functional theory (DFT) calculations.
• Constructed training sets combining heterogeneous DFT exchange-correlation functionals without artificial accuracy hierarchies.
• Compared the multitask framework with existing kernel approaches based on Δ-learning.

Main Results:

• Achieved coupled-cluster (CC)-level accuracy in predictions using multitask surrogates.
• Reduced data generation costs by over an order of magnitude compared to traditional methods.
• Demonstrated that multitask regression can accommodate diverse training set structures, including varying levels of data fidelity.

Conclusions:

• Multitask Gaussian process regression effectively reduces data generation costs for molecular property prediction.
• The approach enables high predictive accuracy by efficiently utilizing heterogeneous and multi-fidelity data.
• This method offers a powerful tool for accelerating computational chemistry research by exploiting existing data sources.