Session: Government Agency Student Posters

Paper Number: 172556

Explainable Machine Learning Models for Real-Time Estimation of Melt Pressure and Temperature in Polymer Extrusion Processes 

Melt pressure and melt temperature are critical indicators of product quality, energy efficiency, and process safety in polypropylene sheet extrusion. However, accurately predicting these parameters in real time remains a significant challenge due to the complex, nonlinear dynamics of the process and the presence of sensor noise and time delays. This study presents a comprehensive and explainable machine learning (ML) framework designed to model these key variables using real-world data collected from an industrial single-screw extrusion system.

The proposed methodology begins with time alignment of process variables through lag estimation using cross-correlation functions, followed by signal smoothing via a fifth-order Butterworth filter. These preprocessing steps reduce high-frequency noise and correct temporal mismatches in the dataset. Informed by domain expertise, a set of engineered features is generated to capture thermomechanical interactions in the extrusion process, enhancing model interpretability and predictive power.

Several supervised ML models are evaluated, including Light Gradient Boosting Machine (LightGBM) and Gradient Boosting Regressor (GBR), with hyperparameters optimized using grid search. On raw downsampled data (0.1 Hz), LightGBM achieves an RMSE of 0.1768 °C and R² of 0.91 for melt temperature, while GBR records an RMSE of 0.0277 MPa and R² of 0.72 for melt pressure. Retraining the models on Butterworth-filtered, lag-adjusted data at 10 Hz yields significant improvements: LightGBM reaches an RMSE of 0.0367 °C and R² of 0.9954, while GBR achieves an RMSE of 0.0032 MPa and R² of 0.996 on the holdout set. These correspond to 79.7% and 85% accuracy improvements, respectively.

Importantly, LightGBM is capable of handling missing (NaN) values natively during both training and inference. This enables continued operation even under partial sensor failure, with large deviations in output predictions serving as indirect indicators of system faults. The low computational overhead of the trained models—sub-microsecond inference times—positions them as viable candidates for integration into real-time control architectures such as edge devices or programmable logic controllers (PLCs).

To ensure transparency, a multi-method explainability suite incorporating SHAP (Shapley Additive Explanations), LIME (Local Interpretable Model-Agnostic Explanations), and PDPs (Partial Dependence Plots) is applied. Analysis reveals that barrel zone temperatures are most influential in predicting melt temperature, while screw speed plays a dominant role in predicting melt pressure.

In conclusion, this study offers a robust, interpretable, and real-time-capable ML framework for modeling critical process parameters in polymer extrusion. The results pave the way for intelligent, closed-loop control and broader industrial adoption of explainable AI in manufacturing environments.

Presenting Author: Mohammad Akram University of New Haven

Presenting Author Biography: Mohammad Basit Akram is a Ph.D. candidate in Engineering and Applied Science Education program at the University of New Haven focusing on mechanical engineering. His research focuses on predictive modeling and advanced machine learning techniques for process optimization in manufacturing systems. With expertise in time series forecasting and explainable AI, he develops data-driven solutions for real-time monitoring and control of polymer extrusion processes. He holds an M.S. from Lehigh University, Pennsylvania and a bachelor’s in engineering from Jadavpur University, India.

Authors:

Mohammad Akram University of New Haven
Ganesh Balasubramanian University of New Haven