On machine learning and visual analysis for quality prediction of film metallization process

Data-driven systems have been increasingly applied in industries to improve process production and pattern analysis. To enhance an industrial vacuum metallization process, we propose advanced machine learning (ML) technologies to extract information, make predictions, and elaborate prescription scenarios. We also implemented visual tools to promote robustness, interpretability, and reliability based on visual interaction between models and operators. The random forest algorithm demonstrated the best classification performance in most analyzed metrics throughout an automatic ML implementation, with 85.4% of accuracy and 0.76 of area under curve (AUC). Media optical density is the most critical product feature for quality analysis with a positive impact on higher values, followed by the warm-up time of ceramic boats, which present better stability to extended warming times. Moreover, specific ranges of operating conditions were identified, such as wire speed and warm-up time, enabling higher values for optical density variables and offering the best conditions for film approval. Finally, visualization techniques allowed us to interpret feature importance, correlation, and patterns that directly interfere with product classification. A product summary enables observing this interference and predicting the probability of approval of a specific product manufactured. The results showed that visual tools and ML algorithms are promising for industrial automation, monitoring, and process improvement. The proposed approach can support analysts and operators in quality analysis and process management.