Machine learning (ML) in science and STEM education: a systematic review

Especially in Machine Learning (ML), the quick development of Artificial Intelligence (AI) has heightened interest in its ability to aid science and STEM education. Still, ML has frequently been used as a technical tool for assessment and analysis inside Science and STEM education, with little integration of how these uses are educationally included and how they affect student learning and scientific thinking. With a particular emphasis on instructional roles, learning results, and persistent problems pertinent to science and STEM teaching and learning, this study offers a Systematic Analysis of empirical research investigating ML usage in science and STEM education. Following the PRISMA framework, a systematic search of the Scopus Database (SD) found 24 peer-reviewed, open-access empirical studies published between 2020 and 2025 investigating ML applications in Science and STEM education settings and K−12. The provided studies were examined in light of STEM discipline, instructional design, ML methods, learning environments, and stated effects on scientific thinking abilities as well as academic achievement. The results showed that ML in STEM education is mostly used as an analytical and diagnostic framework to facilitate automated assessment, learning analytics, misconception detection, and learning progression modeling. Most often used are methods of supervised learning, natural language processing, deep learning, and clustering approaches, especially in physics, integrated science, and data-rich Science and STEM learning contexts. ML solutions usually serve as teacher-facing decision-support systems rather than as learner-centered adaptive tools in several investigations. Learning effects demonstrate that ML most significantly helps Science and STEM education when integrated into pedagogically deliberate teaching approaches. Studies combining curriculum-aligned interventions, ML-driven feedback, and formative assessment reveal gains in scientific reasoning, problem-solving, and conceptual grasp. ML applications, on the other hand, concentrated mostly on prediction correctness or assessment automation, showing minimal direct impact on student success; their advantages are mediated via instructional decision-making rather than by means of student contact with ML systems. The review also points out important obstacles and research gaps: the prevalence of non-experimental designs, scant causal and longitudinal evidence, modeling higher-order Science and STEM thinking challenges, concerns about interpretability and fairness, and not enough focus on teacher professional capacity for ML-supported instruction. Generally, the results showed that although ML has improved assessment and analysis in Science and STEM education, its instructional and transformative potential is still untapped. Future studies in Science and STEM education should give theory-driven, equity-aware, and experimentally proven ML-integrated educational systems priority since they help to develop deeper scientific thought and long-term learning results.