(711b) A Digital Twin Simulator of a Pastillation Process with Applications to Computer Vision-Based Control

Pastillation processes are an important solidification technology that convert a molten product (i.e., polymer melts) into uniform dispersed hemispherical solid forms called pastilles, thus offering a dust-free alternative to traditional mechanical cutting and breaking processes [1-2]. However, some of the common operational challenges of pastillation processes involve the fluctuations in pastille flow rates, sizing, and temperature control, primarily stemming from limited frequency and effectiveness of manual visual inspections of the conveyor belt, as well as lack of real-time quantitative measurements of pastille production rates. To address these limitations, this work presents a digital twin simulation framework that generates realistic thermal image data of the pastillation process. The developed framework facilitates the training of computer vision-based soft sensors, specifically convolutional neural networks [3] (CNN), to predict key process variables such as temperature and flow rate. The output signals produced by the computer vision sensor (called as PaNet – Pastillation Network) enable real-time process monitoring and feedback control [4].

The proposed pastillation conveyor belt simulator [5] models the spatial distribution, cooling, and production rate of the pastilles as they move down the conveyor belt while incorporating random clogging events. The simulation generated realistic thermal image data, mimicking an overhead thermal camera, that was then used to train the computer vision-based sensor (PaNet). PaNet employed CNN architectures (1D and 2D) to predict the instantaneous average flow rate and temperature of the pastilles. Furthermore, the digital twin framework incorporates a closed-loop feedback control system using PaNet and proportional-integral-derivative (PID) controllers tuned using Bayesian optimization [6] to control the temperature and flow rate of pastilles, thus maintaining production efficiency. We observed that the 1D CNN architecture demonstrated superior performance in predicting both temperature and flow rate, with saliency map analysis [7] confirming its effective focus on critical thermal image regions.

The results obtained with the digital twin simulator highlight the potential of using computer vision to process control in pastillation processes. Specifically, we show that the PaNet 1D-CNN model accurately predicted temperature and flow rate, thus enabling effective closed-loop PID control to achieve desired temperature setpoints. Our results aim to pave the way towards the development of new algorithms and tools for computer vision and controls in manufacturing processes.

References:

1. Kim, J.-W. & Ulrich, J. Prediction of degree of deformation and crystallization time of molten droplets in pastillation process. Int J Pharm 257, 205–215 (2003).

2. Chen, C.-C. Continuous production of polyester-poly (ethylene terephthalate) resins in melt-phase and solid-state reactors. Polym Eng Sci 57, 505–519 (2017).

3. Jiang, S. & Zavala, V. M. Convolutional neural nets in chemical engineering: Foundations, computations, and applications. AIChE Journal 67 (2021).

4. Gao, S., Dai, Y., Li, Y., Jiang, Y. & Liu, Y. Augmented flame image soft sensor for combustion oxygen content prediction. Meas Sci Technol 34, 015401 (2023).

5. González, L. D., Pulsipher, J. L., Jiang, S., Soderstrom, T. & Zavala, V. M. A Digital Twin Simulator of a Pastillation Process with Applications to Automatic Control based on Computer Vision. arXiv preprint arXiv:2503.16539 (2025).

6. González, L. D. & Zavala, V. M. New paradigms for exploiting parallel experiments in Bayesian optimization. Comput Chem Eng 170, 108110 (2023).

7. Smilkov, D., Thorat, N., Kim, B., Viégas, F. & Wattenberg, M. SmoothGrad: removing noise by adding noise. arXiv preprint arXiv:1706.03825 (2017).