D2VP: A Visual First Platform for Diagnosing Data Distribution Mismatch Toward Reliable Machine Learning in Materials...

The reliability of machine learning (ML) models in chemistry and materials discovery is often compromised by Data Distribution Mismatch (DDM), a systematic deviation between training data distributions and target chemical spaces manifesting in three dimensions: over-density (redundant sampling of known regions), sparsity (insufficient coverage of the feature space), and extrapolation risk (predictions beyond the training domain). To address this challenge, we propose a data-centric, visual-first methodology that transforms reliability assessment from abstract metrics into evidence-based visual inquiry, operationalized through D2VP (Data Diagnosis & Visualization Platform), an open-source interactive system where "2" denotes the dual capability of diagnosing distribution anomalies and visualizing high-dimensional chemical spaces in a human-in-the-loop environment. Case studies validate the methodology across each DDM dimension. Redundancy-reduced sampling reduces dataset size by over 60% with negligible accuracy loss. Exploration-based sampling guides new experiments; adding just 10% of targeted data reduces mean absolute error by over 80% for models trained on sparse data. The credibility analysis toolkit identifies high-risk extrapolation scenarios, including a case with a training-to-prediction ratio of nearly 1:400. By integrating these capabilities, D2VP transforms abstract reliability metrics into actionable scientific insights, providing a domain-agnostic framework for building trustworthy predictive models across the AI for science landscape.