Building Trustworthy AI for Materials Discovery: From Autonomous Laboratories to Z-scores

Most AI/ML studies in materials discovery fall short on trustworthiness: The research found that published studies achieved a median score of only 5 out of 7 on the GIFTERS trustworthiness framework, indicating that "comprehensive approaches to trustworthiness are rarely reported" in this field.

Different AI approaches have different blind spots: Bayesian studies frequently neglect fair data practices, while non-Bayesian studies most often omit interpretability—suggesting that researchers should be aware of which trustworthiness principles their chosen methods tend to overlook.

The field needs a structured framework for trustworthiness: The GIFTERS framework identifies seven key principles—generalizability, interpretability, fairness, transparency, explainability, robustness, and stability—that the materials discovery community values, providing a concrete checklist for evaluating whether AI models are truly reliable.

Other scientific fields have already solved some of these problems: Healthcare, climate science, and natural language processing have developed methods for ensuring trustworthy AI that could be adapted for materials discovery, rather than the field reinventing solutions from scratch.

Human oversight remains essential: The paper emphasizes that "human-in-the-loop" approaches are necessary to bridge the gap between trustworthiness and uncertainty quantification, meaning scientists should remain actively involved in validating and interpreting AI results rather than treating models as black boxes.