STEM education research has traditionally relied on classroom experiments, student surveys, and performance analytics to understand how learners reason through complex problems. However, these approaches require large student cohorts, raise ethical constraints, and offer limited control over cognitive variability. This study introduces a novel methodology that replaces traditional data collection with synthetic students created using large language model agents orchestrated through a Python-based simulation pipeline. Five learning personas (novice, intermediate, competent, overconfident incorrect, and expert) were defined and used to generate 1,250 reasoning traces in response to a curated set of 50 mixed-difficulty problems with five runs each.
Python automation enabled controlled persona behavior, prompt engineering, iterative sampling, and structured extraction of reasoning steps, misconceptions, confidence estimates, and final answers. Through systematic analysis of these traces, we developed the Universal STEM Misconception Taxonomy, which captures cross-disciplinary cognitive errors such as incorrect rule generalization, symbolic–verbal disconnects, inverse reasoning, premature formula application, dimensional inconsistency, and flawed causal assumptions in physical systems. The synthetic students reproduced well documented STEM misconceptions and exhibited stable, persona-specific error signatures that could be computationally analyzed at scale. The expert persona served as an internal benchmark for validity, enabling automated difficulty profiling and reasoning quality assessment.
This work demonstrates that synthetic learners combined with reproducible Python workflows offer an ethical, scalable, and cost-effective paradigm for STEM education research. The methodology removes the need for human participant recruitment while enabling high-fidelity modeling of diverse problem solving behaviors. The proposed framework supports new applications in STEM education, including curriculum diagnostics, automated distractor generation, pre-testing of instructional materials, and the development of learning analytics. The pipeline is fully platform-independent and can be extended to any STEM discipline, positioning synthetic cognition as a transformative approach for the next generation of STEM education research.
