Challenge
Data quality is degraded by duplicate entries and errors across production, material, and property data. The high number of recipe variations increases the time and cost of experimentation, while the limited availability of labeled data makes it difficult to train robust models. When evaluating new raw materials, deviations from existing data patterns reduce model reliability—requiring continuous feedback from the factory floor to maintain performance.
Approach
Checklist codes were introduced to manage and validate material property data. Recipe variables were consolidated to reduce dimensionality and minimize the number of required experiments and labels. Context-specific test scenarios were designed to identify accuracy variations under different conditions, and a feedback dashboard was developed to collect insights from the production floor in real time—enabling continuous model refinement.
Value Delivered
High-quality data improved the predictive accuracy of the AI model. Label efficiency optimization helped reduce development time and testing costs. The model’s contextual reliability and scope of application were clearly defined, and a feedback-driven improvement loop was established—supporting process optimization and ongoing productivity gains.
