Advancing Sustainable and Equitable STEM Education: A GAN-CNN Integrated Model for Precise Learning Diagnosis and Individualized Instruction

Sustainable and equitable STEM education requires assessment mechanisms that support timely instructional decisions while remaining feasible in resource-constrained classroom environments. Traditional assessments typically report only class-level statistics, limiting teachers’ ability to diagnose individual learning difficulties. This study proposes a classroom-oriented AI-assisted diagnostic framework that integrates generative adversarial networks (GANs) and convolutional neural networks (CNNs) to support learning pattern identification under conditions of severe data scarcity. Student response-behavior data collected through an online testing platform were used to categorize learners into predefined learning behavior types. The GAN was employed to generate locally perturbed samples for stability-oriented data expansion at multiple scales, while the CNN served as a pattern consistency learner operating on the expanded dataset. Rather than aiming for population-level generalization, the framework examines the stability and consistency of learning behavior classification within a single classroom context. Classification results across different expansion scales showed stable performance, with CNN accuracies exceeding 72%. Based on diagnostic outputs, teachers implemented targeted remedial instruction. Case study results show that four out of five remedial interventions exhibited observable improvement. These findings indicate that the proposed framework functions as a proof-of-concept decision-support tool for formative diagnosis and targeted instruction, supporting more equitable learning opportunities, improving instructional efficiency, and contributing to sustainable STEM education aligned with SDG 4.