Optimization and validation of air impingement cleaning parameters for removing nonfat dry milk powder from stainless steel surfaces by machine learning

Air impingement cleaning can remove residues from food contact surfaces without introducing moisture, thereby controlling allergen cross-contact and pathogen cross-contamination in dry food manufacturing environments. The efficiency of air impingement was affected by the following parameters: nozzle diameter (ND), nozzle height (NH), air pressure (AP), water activity (a_w), and residue thickness (THK). In this study, a database from our previous experiment was constructed and employed with machine learning (ML) to determine optimal cleaning conditions (ND, NH/ND, AP, a_w, and THK), which require minimal time to remove nonfat dry milk from the stainless steel (SS) surface and induce the maximum wall shear stress (WSS) on SS surfaces by air impingement. Response surface methodology and 6 different ML models (eXtreme gradient boosting, random forest, gradient boosting regression, backpropagation neural network, support vector regression, and k-nearest neighbor [KNN]) were developed. Among the 7 models, the KNN model demonstrated the best performance in predicting removal time and WSS, with the highest R² and lowest root mean squared error. The optimized conditions predicted by ML models were experimentally validated; the maximum error between the predicted value of the KNN model and the experimental value of air impingement cleaning for removal time and WSS was within 15% and 3%, respectively, which demonstrated the model's accuracy. With further development, the ML models could help optimize and facilitate the practical application of air impingement in dry cleaning.