Predictive Modelling of Silver and Thickness Measurements in Medical Foam Production - Interpretable Machine Learning for Process Understanding

Abstract

Medical foam used in wound-care products must meet quality requirements for properties such as thickness and silver content. This thesis investigates whether routinely collected production data can be used to predict these properties and to support process understanding. The study uses six months of data from medical foam production, including quality-control and laboratory measurements, roll information, and process sensor data from the production equipment. The raw data sources were cleaned, linked to finished production rolls or roll positions, and converted into modelling tables for four responses: silver concentration, Ag [mg/g]; silver area-density, Ag [mg/cm2]; average roll thickness; and thickness log variance, which describes variation in thickness within a roll. Several regression models were compared using time-based cross-validation and a held-out later production period, so that model performance was evaluated in a setting similar to predicting future production. The results depended strongly on the response. Silver concentration was not usefully predictable from the available data, with test R2 = 0.005. Silver area-density showed some predictive signal during the earlier development period, but the models did not generalise to the later test period because silver levels shifted upward. In contrast, average thickness was predicted very accurately from process sensor summaries alone, with the best model reaching test R2 = 0.989. Thickness log variance was harder to predict but still showed useful signal, with the best model reaching test R2 = 0.652. The most stable interpretation results therefore concern thickness. The fitted models identified production contexts and sensor summaries that can guide expert process review, especially for thickness level and foam uniformity. These results should be read as predictive associations, not as causal evidence or direct control rules.