Identifying and embedding transferability in data-driven representations of chemical space

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Transferability is crucial for model generalization across scientific disciplines.
• Complex machine learning models can obscure how data transferability is achieved or missed.
• Identifying and embedding transferable training data are key challenges in data-driven scientific modeling.

Purpose of the Study:

• To develop methods for identifying and embedding transferable training data for ab initio chemical modeling.
• To address the challenge of understanding and improving data transferability in complex machine learning models.
• To enhance the generalization capabilities of data-driven models in chemistry.

Main Methods:

• Introduction of a Transferability Assessment Tool (TAT) for ab initio chemical modeling.
• Demonstration of TAT on a controllable data-driven model for developing density functional approximations (DFAs).
• Analysis of human intuition's impact on training data curation and its chemical biases.

Main Results:

• Human intuition in data curation can introduce chemical biases, hindering the transferability of data-driven DFAs.
• The TAT successfully identified key factors influencing data transferability.
• Three transferability principles were motivated, including the concept of transferable diversity.

Conclusions:

• The developed TAT and identified transferability principles offer strategies to improve data curation for machine learning in chemistry.
• Embedding transferable diversity into training data can enhance the generalization of general-purpose machine learning models.
• This work provides a framework for creating more robust and transferable data-driven models in scientific research.