Transfer learning for a foundational chemistry model.

Area of Science:

• Chemistry
• Machine Learning
• Computational Chemistry

Background:

• Data-driven chemistry relies on large, accurate datasets, which are often challenging to obtain.
• Generating large-scale chemical data is laborious, hindering machine learning applications.
• Existing transfer learning methods require expert task selection and are case-specific.

Purpose of the Study:

• To develop a machine learning framework for accurate chemistry-relevant predictions with limited data.
• To establish a generalizable approach that overcomes common data acquisition challenges in chemistry.
• To demonstrate the efficacy of a foundational model combined with task-specific fine-tuning.

Main Methods:

• Trained a chemical "foundational model" on approximately 1 million experimental organic crystal structures.
• Stacked a task-specific module atop the foundational model.
• Fine-tuned the combined model for various prediction tasks.

Main Results:

• Achieved state-of-the-art performance on diverse prediction tasks.
• Demonstrated accurate predictions for toxicity, yield, and odor.
• Showcased the framework's capability in low-data scenarios.

Conclusions:

• The developed machine learning framework effectively addresses data scarcity in chemistry.
• Foundational models offer a powerful strategy for generalizable chemical predictions.
• This approach advances data-driven chemistry by enabling predictions with limited datasets.