Uncertainty Characterization of Foundation Models for Reliable Applications in Materials and Chemistry

The advent of foundation models has revolutionized the field of materials and chemistry by enabling the rapid characterization of new materials with minimal data. However, these models are not immune to uncertainty, which can significantly impact their reliability and limit their adoption in industrial applications. It is crucial to develop and integrate uncertainty quantification methods that account for the limitations of the datasets used to train the models, including their coverage and biases, as well as the specific requirements of the application domain. This is essential to ensure that the models' predictions are trustworthy and actionable in real-world scenarios. This work highlights the importance of uncertainty characterization in foundation models and presents a framework for combining multiple quantification methods that consider dataset coverage and domain-specific factors to provide a comprehensive understanding of uncertainty in various use cases, including energy storage and PFAS. Furthermore, we emphasize the need for tailored visual analytics tools to effectively communicate uncertainty to domain experts, enabling informed decision-making and facilitating the widespread adoption of AI in materials and chemistry.