Developing Deep Learning in Science Through an Adaptive AI-Based STEM Instructional Program: Evidence From Sixth-Grade Classrooms

Synthesis: This concurrent mixed-methods, cluster-randomized posttest-only pilot (N = 30 sixth-graders in a female-only primary school) evaluated an adaptive AI-based STEM instructional program (personalized content, rule-based mastery, real-time feedback) for developing deep learning in science. Quantitative results favored the experimental group across explanation, interpretation, application, and idea generation (large within-sample effect sizes), corroborated by teacher interviews — though the two-intact-classroom design warrants cautious interpretation.

Key Findings

1. Adaptive AI-based STEM program supports deep learning. The program — personalizing content, adjusting task difficulty, and providing real-time feedback over eight weeks, with rule-based mastery governing adaptive pathways — produced statistically significant differences favoring the experimental group across all deep-learning dimensions.
2. Large within-sample effect sizes. Explanation, interpretation, application, and idea generation showed substantial rank separation, though classroom-level confounding and baseline differences cannot be fully ruled out in a two-cluster design.
3. Teacher corroboration. Interviews with three science teachers reported perceived improvements in analytical reasoning, conceptual integration, inquiry-based exploration, and creative scientific thinking; simulations, adaptive feedback, and hands-on STEM activities supported conceptual understanding and engagement.
4. Rule-based mastery + ML monitoring. Adaptive learning pathways were governed by rule-based mastery mechanisms, with Machine Learning components used solely for performance monitoring rather than direct trajectory modification.

What this means for practice

• Instructors. Pair adaptive AI pathways with inquiry-oriented STEM tasks rather than running them as standalone content delivery: the program's gains covered explanation, interpretation, application, and idea generation, and the three interviewed teachers tied conceptual understanding to simulations, adaptive feedback, and hands-on activities working together.
• Designers. Keep mastery decisions rule-based and confine machine learning to performance monitoring, as this program did, so adaptive trajectories stay auditable and human oversight is preserved.
• Instructors. Target deep-learning dimensions explicitly — explanation, interpretation, real-world application, and idea generation — instead of treating adaptive personalization as a route to surface content coverage.
• Designers. Treat this program as a pilot template to validate locally before scaling: the authors present it as context-specific evidence from one school and recommend larger multi-site replication with stronger cluster-level controls.

Limitations

• N = 30 sixth-graders in two intact classes with one classroom per condition (cluster-randomized posttest-only), so classroom-level confounding and baseline differences cannot be ruled out and student-level estimates are treated as exploratory.
• The study ran in a single female-only primary school in one context over eight weeks, which the authors state should be validated through larger, multi-site studies.
• There was no pretest, so group equivalence was never established; deep learning was measured only at posttest with median (IQR) summaries and Mann–Whitney U tests in a two-cluster design.
• Qualitative corroboration comes from interviews with three science teachers — a small, non-independent source relative to the student outcome data.

Citation

Bin Bakheet, T., Alamri, H., & Alshaya, F. (2026). Developing deep learning in science through an adaptive AI-based STEM instructional program: evidence from sixth-grade classrooms. International Journal of STEM Education, 13, 35.