Kunststoff und Innovation

Polylactic Acid (PLA) has become a cornerstone of sustainable polymer solutions, widely used in packaging, textiles, 3D printing, and medical devices. Its popularity stems not only from its biodegradability but also from its versatility and performance. One of the most critical parameters that determines the quality of PLA is its molecular weight. Molecular weight directly influences mechanical properties, thermal stability, processability, and degradation rate. In medical-grade PLA, which must meet stringent FDA regulations, maintaining consistent molecular weight during extrusion processing is particularly crucial.

Traditional methods for monitoring molecular weight rely on laboratory-based testing, which can be time-consuming and limited in providing real-time feedback. To address this challenge, researchers at Atlantic Technological University, Ireland, explored the potential of machine learning algorithms to predict PLA molecular weight directly from in-line process data. Their study combined Artificial Bee Colony (ABC) optimization, a nature-inspired algorithm based on honeybee foraging behavior, with two supervised learning techniques: Artificial Neural Networks (ANNs) and Adaptive Neuro-Fuzzy Inference System (ANFIS). This integration not only improved prediction accuracy but also enhanced interpretability by identifying the most influential features among hundreds of inputs.

During the PLA extrusion process, Near-Infrared (NIR) spectroscopy was used alongside process data, such as melt temperature, pressure, and machine settings. This setup produced a dataset with 63 observations and 512 input features, reflecting the high dimensionality and complexity inherent to polymer processing data. While such datasets are rich in information, they pose challenges for standard machine learning approaches. Redundant and irrelevant features can increase computational effort, obscure model interpretability, and even reduce prediction accuracy.

To overcome these challenges, the ABC algorithm was employed for feature selection. By mimicking the foraging behavior of honeybee swarms, ABC explores the search space efficiently, identifying the most informative features while minimizing the number of inputs. This approach was applied alongside ANFIS and ANN regression models to determine the minimal set of inputs required to accurately predict molecular weight. The results were striking. Using the ABC-ANFIS combination, the researchers reduced the original dataset to just four critical features: three specific NIR wavenumbers (6158, 6310, 6349 cm⁻¹) and melt temperature. Despite this significant reduction, the model achieved a root mean square error (RMSE) of only 282 Da and an R² of approximately 0.96 during five-fold cross-validation, demonstrating both high accuracy and robustness.

The interpretability of this approach is particularly valuable for polymer processing. By identifying the specific NIR signals and process parameters that most influence molecular weight, operators and engineers can monitor and adjust processing conditions in real time. Compared to previous methods, such as Recursive Feature Elimination with random forests, which required nine features for prediction, the ABC-ANFIS method offers a simpler and more transparent model without compromising accuracy.

For Plastrans Technologies, these advances in interpretable machine learning are highly relevant. Applying such techniques to PLA and other bio-based polymers can support real-time quality monitoring, optimized processing, and consistent product performance. By focusing on the most significant features of a process, manufacturers can reduce material waste, enhance efficiency, and maintain the high standards required for medical, packaging, and consumer applications.

This research highlights the power of combining advanced optimization algorithms with supervised learning to handle complex, high-dimensional datasets. It offers a pathway toward smarter, more sustainable polymer production, enabling a deeper understanding of material behavior and more reliable control over product quality. For industries that depend on PLA and other biopolymers, this represents a step forward in achieving both operational excellence and environmental responsibility.

Source:

Masoumi, A. P., Creedon, L., Ghosh, R., Munir, N., McMorrow, R., & McAfee, M. (2024, June). Interpretable machine-learning for predicting molecular weight of PLA based on artificial bee colony optimization algorithm and adaptive neurofuzzy inference system. In 2024 35th Irish Signals and Systems Conference (ISSC) (pp. 1-6). IEEE.