NUS Explainable-AI Framework for Scientific Discovery 2026

The National University of Singapore’s College of Design and Engineering (NUS CDE) reported a Perspective proposing a framework for using explainable artificial intelligence (XAI) in scientific discovery and high-stakes applications. The Perspective was co-authored by NUS Assistant Professor Gianmarco Mengaldo and published in Nature Communications on 6 August 2026; NUS published its Singapore research feature on 2 September 2026 (NUS CDE, 2 September 2026).

From prediction to testable hypothesis

The proposed workflow treats an AI explanation as a lead for investigation rather than as a scientific explanation that can be accepted on appearance alone. Researchers would identify which inputs or patterns influenced a model’s prediction, translate that clue into a hypothesis, and then test it through experiments, simulations or established scientific principles. NUS illustrates the idea with AI analysis of turbulent airflow: an explanation can identify a feature worth investigating, while further testing determines whether it reflects a genuine physical process or merely a recurring pattern in the training data (NUS CDE, 2 September 2026).

Discovery, optimisation and certification

NUS describes three linked uses for the approach. Explainability can support discovery by surfacing patterns that suggest how a system functions; tested insights can support optimisation by improving a design or control strategy; and records of what influenced a model can contribute to certification and reliability assessment. In a reported turbulence example, XAI identified influential airflow patterns and helped researchers train a second AI system to reduce drag more effectively than a system trained only on the final drag outcome (NUS CDE, 2 September 2026).

High-stakes boundary

For healthcare, aviation, energy, infrastructure and autonomous systems, the feature says evaluators would need to examine both model accuracy and whether an explanation faithfully reflects what influenced the model, including under unfamiliar conditions. Explanations may strengthen safety and regulatory assessments, but NUS explicitly presents them as supplementary evidence rather than replacements for existing testing, certification or regulatory processes. The article describes a proposed framework and ongoing exploration, not a deployed regulatory standard or a claim that the method has already validated every high-stakes AI system (NUS CDE, 2 September 2026).