Optimizing Energy-Intensive Heating Processes in Manufacturing Using Bayesian Belief Networks: A Case Study in Thermoforming

Iman Jalilvand, Niloofar Akbarian-Saravi, Bryn Crawford, Abbas S. Milani, R. Bhushan Gopaluni2025 · DOI

TL;DR

Researchers at the University of British Columbia have developed a Bayesian Belief Network (BBN) framework to optimize thermoforming—a notoriously complex and energy-hungry manufacturing process. By moving away from "black-box" AI and toward an interpretable, probabilistic model, they’ve created a tool that helps engineers balance high product quality with low energy consumption, even when data is scarce.

What Problem Does This Paper Address?

Thermoforming is the process of heating a plastic sheet until it’s pliable and then stretching it over a mold to create 3D shapes (think of everything from car dashboards to food packaging). While it sounds simple, it is a nightmare to control. The process is highly non-linear and sensitive to "noise"—slight changes in ambient factory temperature, airflow, or material thickness can lead to warping, thinning, or structural defects.

Traditionally, engineers have relied on trial-and-error or "black-box" machine learning models like Artificial Neural Networks (ANNs). However, ANNs require massive datasets that many factories don't have, and they offer zero explanation for why they make a certain prediction. This paper tackles the need for a "transparent" decision-support tool that can handle uncertainty and integrate human expertise with experimental data.

Methods

The authors utilized Bayesian Belief Networks (BBNs). Unlike standard neural networks, a BBN is a graphical model that represents variables and their conditional dependencies using a "Directed Acyclic Graph" (essentially a flowchart of cause-and-effect).

Key methodological highlights include: * Hierarchical Structure: The model is organized into three layers: Inputs (Power, Initial Temp, etc.), Intermediates (Error, Settling Time, Stability), and Outputs (Product Quality, Energy Usage). * Expert-Guided CPTs: Instead of training the model on millions of data points, the authors used Conditional Probability Tables (CPTs) built via a weighted arithmetic mean. They mapped these scores to probabilities using a sigmoid-shaped logistic function, allowing the model to reflect real-world engineering logic. * Reverse Inference: One of the most powerful features of this BBN is the ability to work backward. An engineer can set the desired outcome (e.g., "100% High Quality") and the model will infer the most likely input configurations needed to get there.

Datasets

The study did not rely on a single massive public dataset. Instead, it integrated:

  • Experimental Observations: Data from a real-world thermoforming setup.
  • Computational Modeling: Results from prior physics-based simulations.
  • Expert Insights: Domain knowledge used to weight the importance of different variables.
  • Discretized Physical Ranges: The model used specific ranges (e.g., Power from 0–500W, Initial Temp from 25–100°C) derived from a prior 2025 study by the same lead author.

Key Insights and Ideas

  • The "Preheat" Advantage: The sensitivity analysis revealed that Initial Temperature is the single most influential factor. Setting it to "High" significantly improved every single outcome metric, including energy efficiency. This suggests that investing in better preheating is more effective than cranking up the heater power during the forming stage.
  • The Power Paradox: High power inputs often led to process instability and thermal "overshoot." The model showed that "moderate" power, combined with high material conductivity, actually produces more consistent results.
  • Transparency over Complexity: The BBN successfully identified that "Error" and "Process Stability" account for over 70% of the variation in product quality. This gives manufacturers a clear "hit list" of what to monitor on the factory floor.

Key Results

The researchers tested the model through two primary scenarios:

  1. Case 1 (Targeting High Quality): When the model was forced to produce high-quality parts, it revealed that the most probable path involved Low Power (77.1%) and Low HTC (75.2%). This counter-intuitive result shows that aggressive heating/cooling actually hurts quality.
  2. Case 2 (Targeting Low Energy): The model found that energy savings could be achieved without sacrificing quality, provided the material had high thermal conductivity and a favorable initial temperature.
  3. Sensitivity Heatmap: Fixing the Initial Temperature to "High" resulted in a 68.9% probability of Low Error and a 72.8% probability of Low Energy Use, proving the model's ability to find "win-win" configurations.

Limitations and Future Work

While the model is a significant step forward, the authors acknowledge a few hurdles:

  • Manual CPT Construction: Currently, the "logic" of the network is largely defined by experts. Future work aims to automate this process using data-driven learning to make the model even more adaptive.
  • Utility-Based Optimization: The researchers plan to extend the framework to include "utility nodes," which would allow the model to make specific economic recommendations (e.g., "Is the extra energy cost worth the 2% increase in quality?").
  • Material Integration: Integrating more complex multi-criteria material selection methods would allow the model to recommend different plastic composites for specific shapes.

Summary updated

Abstract

This study investigates the application of Bayesian Belief Networks (BBNs) to optimize thermoforming, an energy-intensive manufacturing process that rapidly shapes thermoplastic composite sheets into precise three-dimensional forms. Thermoforming involves complex interactions among process parameters, presenting ongoing challenges in achieving optimal energy efficiency, temperature stability, and consistent product quality. Further, the process can become highly non-linear with multiple sources of noise, due to the nature of the factory and logistical factors, adding further challenges to effectively predicting the process outcomes. To address this challenge, in this case study a Bayesian Belief Network was developed to capture the inherent uncertainties and intricate relationships among critical process variables. This probabilistic model integrated experimental observations, computational modeling results, and expert-derived insights, enabling a robust and adaptive decision-making support for the process designers. Each of the select key process parameters such as convective Heat Transfer Coefficient (HTC), and heating power, were prioritized based on their influence on the process performance indicators such as the end product quality, energy usage, temperature distribution error, settling time, and stability. Results demonstrated that the BBN framework provides an effective interactive decision support tool capable of continuous model refinements through updating of probabilities as new data became available, while achieving enhanced energy efficiency and process control, thereby also reducing the operational cost and material wastage.
BibTeX
@misc{jalilvand2025optimizing,
  author = {Iman Jalilvand and Niloofar Akbarian-Saravi and Bryn Crawford and Abbas S. Milani and R. Bhushan Gopaluni},
  title = {{Optimizing Energy-Intensive Heating Processes in Manufacturing Using Bayesian Belief Networks: A Case Study in Thermoforming}},
  year = {2025},
  doi = {10.46254/eu08.20250283},
}
Figure