Diverse Explanations From Data-Driven and Domain-Driven Perspectives in the Physical Sciences

Multiple valid explanations can exist for the same ML prediction: The paper demonstrates through three case studies in materials science that different explanation methods and models can produce varying—yet scientifically valid—interpretations of why an AI system made a particular decision, meaning there is no single "correct" explanation.

Scientists must remain in control of AI interpretation: The authors emphasize that domain experts should actively guide how ML explanations are interpreted rather than passively accepting data-driven insights, balancing algorithmic outputs with their scientific knowledge to ensure results meet specific research needs.

Different stakeholders may need different explanations: The paper shows that the level of detail and type of explanation required varies depending on who is using the AI system (e.g., researchers designing experiments vs. those validating discoveries), highlighting that one-size-fits-all explanations are insufficient for responsible AI deployment in science.

Explanation diversity can lead to scientific inconsistency if unmanaged: Without careful consideration of why explanations differ, researchers risk drawing conflicting conclusions from the same ML model, potentially undermining the scientific value of AI-assisted discovery and eroding trust in these tools.

Responsible AI in science requires transparency about limitations: The paper advocates for scientists to understand not just what an AI explains, but how and why different explanation methods work, enabling them to make informed decisions about when and how to trust AI recommendations in their research.