⚡ electrical engineering

Model-Guided Local Bayesian Optimization for Tuning of Interpretable Controllers in Injection Molding

This paper proposes a model-guided local Bayesian optimization framework that leverages a physics-inspired neural mixture-of-local-experts model augmented with a residual Gaussian process to efficiently and safely tune interpretable controllers for injection molding, achieving performance comparable to global optimization while mitigating high-cost risks during the tuning process.

Original authors: Jens Ahlers, Robert Göllinger, Xu Chen, Heike Vallery, Sebastian Stemmler

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jens Ahlers, Robert Göllinger, Xu Chen, Heike Vallery, Sebastian Stemmler

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master chef trying to perfect a recipe for a very delicate soufflé. The goal is to get the texture and rise (the "quality") just right. In the world of plastic manufacturing, this "soufflé" is a plastic part, and the "kitchen" is an injection molding machine.

Here is a simple breakdown of what the researchers in this paper did, using that kitchen analogy.

The Problem: The "Black Box" Chef

Currently, advanced computer programs can control these machines very well. However, they are like a black box chef: they know how to make the perfect soufflé, but they won't tell you why they added a pinch of salt or turned the heat up.

The Issue: Factory workers don't trust a chef they can't understand. If the machine makes a mistake, they can't fix it because they don't know how the "chef" thinks.
The Challenge: Every time you change the shape of the mold (the pan) or the type of plastic (the batter), the recipe needs to be completely rewritten. Doing this by trial and error is expensive because if you mess up, you ruin the plastic (scrap parts) and might even break the machine.

The Solution: The "Transparent" Chef with a Safety Net

The authors propose a new way to tune the machine's controller (the "chef") that is interpretable (easy to understand) and safe.

They created a method called Model-Guided Local Bayesian Optimization (MGLBO). Here is how it works, broken down into three parts:

1. The "Physics-Inspired" Simulator (The Recipe Book)

Instead of just guessing, the system uses a digital twin—a computer simulation of the machine.

The Analogy: Think of this as a highly detailed recipe book that understands the physics of the batter. It knows that if you pour the batter too fast, it splashes; if you wait too long, it cools down.
The Twist: This isn't just a static book. It's a "Mixture of Local Experts." Imagine a team of three specialists: one expert for when the mold is empty, one for when it's half-full, and one for when it's full. They work together to predict exactly what will happen next.

2. The "Safety Net" (The Taste Tester)

Even the best recipe book isn't perfect. Real machines have quirks.

The Analogy: The system adds a "Taste Tester" (a Gaussian Process model) that compares the recipe book's predictions with what actually happens in the real kitchen.
How it helps: If the recipe book says "add 5 grams of sugar" but the real machine needs 6, the Taste Tester learns that difference and corrects the prediction for the next time. This keeps the system accurate without needing thousands of failed attempts.

3. The "Local Search" (The Careful Taster)

This is the most important part. Traditional methods try to taste every possible combination of ingredients in the whole world (Global Optimization). This is dangerous because you might accidentally add a cup of salt instead of a pinch, ruining the batch.

The Analogy: The MGLBO method is like a cautious chef who only tweaks the recipe slightly from the last successful batch.
The Trust Region: The chef stays in a "safe zone" (a small neighborhood of parameters). If a small tweak makes the soufflé better, the chef expands the safe zone slightly. If a tweak makes it worse, the chef shrinks the zone and tries a different small tweak.
Risk Awareness: The system is programmed to be "risk-averse." It won't take a huge gamble that might break the machine, even if there's a small chance it could be perfect. It prioritizes not making a mess over finding a theoretical "perfect" solution that might be dangerous.

The Results: What Happened in the Kitchen?

The researchers tested this method in a computer simulation against a standard "brute force" method (Vanilla Bayesian Optimization). They tested three types of "chefs" (controllers):

Simple Chef (P-Controller): Just one knob to turn.
Smart Chef (G-PI): A few knobs that change based on how full the mold is.
Complex Chef (RBF): A very flexible chef with many knobs (30 parameters) that can learn complex curves.

The Findings:

Speed: The new method (MGLBO) found good settings just as fast, or faster, than the standard method, even for the complex chef.
Safety: The standard method often made huge mistakes (high costs) while searching for the perfect setting. The new method rarely made big mistakes. It kept the "cost" (risk of scrap or damage) low throughout the process.
Quality: The final settings found by the new method were just as good as, or better than, the standard method.
Interpretability: Because the controllers were designed to be simple and transparent, the factory workers can actually look at the final settings and understand why the machine is doing what it's doing.

The Bottom Line

This paper presents a way to automatically tune plastic-making machines that is safe, fast, and understandable. It uses a smart computer model to predict the future, a "taste tester" to correct mistakes, and a cautious strategy that only makes small, safe changes. This allows factory workers to trust the machine again, knowing it won't suddenly break the equipment or ruin the product while it learns how to do its job better.

Technical Summary: Model-Guided Local Bayesian Optimization for Tuning of Interpretable Controllers in Injection Molding

Problem Statement
Injection molding (IM) is a critical industrial process where cavity pressure serves as a key determinant of part quality. While advanced model-based control methods can effectively regulate cavity pressure, their industrial adoption is hindered by two primary challenges:

Interpretability: Sophisticated control laws often behave as "black boxes," making it difficult for machine operators to verify outputs and leading to mistrust.
Optimization Complexity: The diversity of mold geometries and material properties makes obtaining accurate first-principles models difficult. Furthermore, optimizing controllers via trial-and-error on physical machines is risky, as it can produce scrap parts or damage equipment.

Existing approaches, such as Vanilla Bayesian Optimization (BO), struggle with high-dimensional input spaces and lack inherent risk awareness during the tuning process.

Methodology
The authors propose Model-Guided Local Bayesian Optimization (MGLBO), a framework designed to automatically tune interpretable controller parameters during the manufacturing process. The approach balances data efficiency with risk awareness by combining a physics-inspired dynamic model with a data-driven correction term.

Process Modeling: The core of the method is a Physics-Inspired Neural Mixture-of-Local-Experts model.
- Drive Model: A Residual Network (ResNet) approximates the drive dynamics (screw velocity and volume) using explicit Euler integration.
- Pressure Model: The pressure dynamics are divided into three phases based on cavity fill level (ambient, moderate increase, rapid increase). These are modeled by three separate "Local Expert" submodels (ResNets) that smoothly transition via a gating layer. The submodels are constrained by physics-inspired equations to ensure physically conforming pressure build-up.
- Training: Models are trained using the Levenberg-Marquardt algorithm on data from the previous $n_C$ cycles, with regularization applied to prevent overfitting to time variables.
Composite Cost Function: The optimization objective is a composite cost function $J_{total}(\theta_C)$ :
$J_{total}(\theta_C) = J_M(\theta_C) + \tilde{J}(\theta_C)$
- $J_M$ : The simulated closed-loop cost derived from the dynamic model.
- $\tilde{J}$ : A Gaussian Process (GP) regression model that learns the residual (mismatch) between the simulated costs and real-world plant observations.
Local Bayesian Optimization:
- The algorithm employs a Trust-Region strategy (based on TurBO-1) rather than global exploration. It defines a trust region $S$ around the current best parameter vector $\theta^*_C$ .
- An acquisition function balances predicted performance against uncertainty: $\alpha(\theta_C) = E(J_{total}) + \kappa\sqrt{Var(J_{total})}$ . The parameter $\kappa$ allows for risk-aware tuning; higher values result in more cautious optimization.
- If a new parameter set yields lower costs, the trust region expands; if performance degrades, the region shrinks. This ensures the algorithm focuses on local optima and mitigates high-cost excursions.
Controllers Tested: The method was evaluated on three interpretable controllers with varying parameter counts (1 to 30):
1. Proportional (P) controller.
2. Gain-Scheduled Proportional Integral (G-PI) controller.
3. Radial-Basis-Function (RBF) controller.

Key Contributions

Novel Optimization Framework: This work presents the first IM process-controller-optimization approach that leverages a dynamic process model combined with plant observations to achieve both data efficiency and risk awareness.
Hybrid Modeling: The integration of a physics-inspired neural mixture model with a GP residual corrector allows for accurate cost prediction without requiring perfect first-principles models.
Risk-Aware Tuning: By utilizing a local trust-region approach and a risk-sensitive acquisition function, the method minimizes the likelihood of generating scrap parts or damaging machinery during the tuning process.

Results
The MGLBO algorithm was benchmarked against Vanilla BO in simulation using 50 independent runs per controller type over 20 optimization iterations.

Performance: MGLBO identified controller parameters yielding minimum costs ( $J_{min}$ ) comparable to or lower than those of global Vanilla BO.
Safety: MGLBO significantly outperformed Vanilla BO in terms of safety metrics. The maximum cost ( $J_{max}$ ) observed during MGLBO tuning remained close to the initial costs, whereas Vanilla BO exhibited substantial "cumulative worsening" ( $J_W$ ), indicating frequent performance deterioration and higher risk.
Convergence: For controllers with higher parameter counts (e.g., RBF), Vanilla BO failed to identify better parameters than the initial guess within 20 iterations, while MGLBO successfully converged.
Execution Time: The mean execution time for the RBF controller optimization was approximately 9.4 seconds per step on standard hardware, which fits within the cooling phase of the IM cycle.

Significance and Claims
The paper claims that MGLBO offers a viable path toward the industrial adoption of advanced control in injection molding by addressing the "black box" issue through interpretable control laws and the "risk" issue through local, model-guided optimization. The authors assert that global optimization methods may be unsuitable for this specific application due to the potential for machine damage from poor parameterization, making the local, risk-aware approach of MGLBO particularly well-suited.

The work concludes that while the computational effort is higher than Vanilla BO due to model retraining, it is feasible within the cooling phase of the cycle. Future work is identified as testing the algorithm on real hardware and extending the model to handle hydraulic drives and dynamic transition points between local experts.