Abstract
In an era where science education increasingly values inquiry competencies over rote outcomes, understanding students’ strategic and non-strategic engagement during digital assessments has become critical. This study applies contrastive representation learning to TIMSS 2023 Grade 8 Earth Science process data from five countries to uncover latent pathways of inquiry behaviour. Real-time digital traces—including screen visit frequency, revisit patterns and time-on-task—were embedded into a structured latent space, revealing two distinct clusters that map onto strategic and non-strategic inquiry profiles when interpreted through SRL theory. K-means clustering and UMAP visualization confirmed a strong alignment between behavioural profiles and Earth Science achievement outcomes. Cross-national analyses demonstrated that strategic and non-strategic engagement structures were highly generalizable across diverse education systems, despite contextual differences. Key behavioural indicators differentiating high- and low-performing profiles were identified, offering new insights into metacognitive regulation during inquiry tasks. By integrating contrastive learning, clustering validation, interpretability and ethical considerations, this study advances process-driven learning analytics for science education. The findings support the development of adaptive assessments, real-time scaffolding tools and culturally responsive instructional strategies, providing a blueprint for how AI-driven methods can enhance inquiry-based learning while safeguarding equity and learner agency in global science education contexts.