Rapid iterative design of reconfigurable smart products via Human–AI collaboration and digital twin feedback

INTRODUCTION: Parametric CAD models for reconfigurable smart products require repeated modification and virtual verification. Existing CAD generation, human–AI co-design, and simulation-assisted validation methods often separate feedback analysis, constraint checking, and CAD editing, limiting the direct use of feedback for executable refinement. OBJECTIVES: This study develops a closed-loop CAD refinement framework that converts design feedback and digital twin–based simulation or constraint feedback into coordinated edit actions while reducing constraint violations, ineffective revisions, and oscillatory updates. METHODS: A conflict-aware dual-feedback framework is proposed. Design feedback is translated into target parameters, local regions, modification directions, and command-level edit conditions. Simulation and constraint results are written back as violation-aware correction cues. When feedback sources conflict, active constraints, historical modification states, and edit magnitudes are jointly considered to generate controlled parameter and operation-sequence updates. The framework is evaluated on feedback-driven benchmarks constructed from the public DeepCAD and SketchGraphs datasets. RESULTS: On the constructed DeepCAD benchmark, the proposed method achieved a constraint satisfaction rate of 90.6% and an average iteration count of 3.43. On the constructed SketchGraphs benchmark, it obtained 89.4% and 3.61, respectively. It outperformed the strongest comparison methods in constraint satisfaction and iterative efficiency while maintaining competitive CAD validity and edit-action accuracy. Statistical and ablation analyses confirmed the contributions of feedback translation, correction write-back, and conflict-aware coordination. CONCLUSION: The results demonstrate the effectiveness of the proposed closed-loop refinement framework under the constructed feedback-driven benchmark settings, particularly in improving constraint satisfaction, reducing refinement rounds, and maintaining stable optimization under conflicting feedback.