A data-efficient foundation model for porous materials based on expert-guided supervised learning

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Foundation models in materials science demand extensive, costly training data.
• Existing physical knowledge (e.g., molecular force fields) is underutilized in current models.

Purpose of the Study:

• To demonstrate how expert knowledge can supervise pre-training, reducing data requirements for foundation models.
• To develop a foundation model for porous materials that effectively integrates structural and energetic information.

Main Methods:

• Developed potential energy surface (PES) basis functions as unified descriptors for guest-host interactions.
• Designed a multi-modal architecture to fuse material structure and PES information.
• Employed pre-training focused on comprehensive geometric feature learning across spatial scales.

Main Results:

• Introduced SpbNet, a foundation model for porous materials trained under limited data conditions.
• SpbNet achieved superior performance on over 50 downstream tasks (adsorption, separation, intrinsic properties) compared to models trained on 20x larger datasets.
• Demonstrated significant error reduction (over 20%) and strong generalization across diverse porous materials (MOFs, COFs, zeolites).

Conclusions:

• Expert knowledge supervision is a viable strategy to drastically reduce data requirements for foundation models in materials science.
• SpbNet offers a powerful and data-efficient approach for predicting properties of porous materials.
• The model's generalization capabilities highlight its potential for broad applications in materials discovery and design.