Hammett-Inspired Product Baseline for Data-Efficient Δ-ML in Chemical Space

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning accelerates molecular and materials design but requires extensive training data.
• High-quality data acquisition is costly and time-consuming.
• Utilizing low-complexity baseline models is crucial for data-efficient learning strategies like Δ-learning.

Purpose of the Study:

• To introduce a novel, data-efficient baseline model for machine learning in chemical compound space.
• To generalize the empirical Hammett equation for broad applicability in molecular and materials design.
• To demonstrate the effectiveness of the proposed model in reducing data requirements for machine learning.

Main Methods:

• Development of a generic coarse-graining Hammett-inspired product (HIP) Ansatz.
• Generalization of the empirical Hammett equation to arbitrary chemical systems and properties.
• Calibration of the HIP Ansatz on various chemical property prediction tasks.

Main Results:

• The HIP Ansatz provides a computationally inexpensive and effective baseline model.
• Demonstrated applicability across diverse chemical properties, including solvation energies, formation energies, adsorption energies, HOMO-LUMO gaps, reaction activation energies, and binding energies.
• HIP serves as a superior baseline for Δ-machine learning, enhancing data efficiency compared to domain-specific models.

Conclusions:

• The Hammett-inspired product (HIP) Ansatz is a versatile and data-efficient baseline model for machine learning applications in chemistry and materials science.
• HIP significantly reduces the data burden associated with machine learning model training.
• This approach offers a promising strategy for accelerating molecular and materials discovery through cost-effective data utilization.