Abstract
In advanced manufacturing, configuring process parameters for technologies such as 3D printing and plasma spraying often involves navigating complex, black-box landscapes where testing is prohibitively expensive or destructive. Standard optimization techniques frequently struggle with the "cost of failure"—sampling parameters that lead to infeasible or damaged parts. This paper introduces a novel Bayesian Optimization (BO) framework that utilizes a switching acquisition function to prioritize feasibility before aggressively pursuing performance gains. Our results demonstrate that this "safety-first" approach achieves optimal configurations with up to 43% fewer trials and produces 2.25 times more feasible samples than industry-standard methods.
1. Introduction
The pursuit of precision in advanced manufacturing—ranging from Atmospheric Plasma Spraying (APS) to Fused Deposition Modeling (FDM)—is often hindered by the high dimensionality and opacity of process parameters. Engineers must balance multiple interconnected outputs, such as material hardness, porosity, and surface finish, against the operational costs of production. Traditional "Design of Experiments" (DoE) approaches are often too resource-intensive, while standard Bayesian Optimization (BO) algorithms tend to be overly "greedy," frequently suggesting parameters that result in failed parts or machine damage.
This research addresses the need for a sample-efficient, status-aware optimization framework that respects the physical and economic constraints of the manufacturing floor.
2. Methodology
The core of the proposed framework is built upon Gaussian Processes (GPs), which are used to model the unknown constraints of the manufacturing process. The framework introduces a two-stage acquisition procedure designed to navigate the trade-off between exploration of the feasible region and exploitation of performance improvements.
2.1 The Switching Acquisition Function
The algorithm employs two distinct acquisition functions:
- Feasible Improvement Probability (FIP): This is a conservative metric used when the algorithm has not yet identified a reliable feasible region. It prioritizes finding any point that satisfies the process constraints, regardless of the potential improvement in the objective function.
- High FIP Improvement (HFI): Once a feasible point is established, the algorithm switches to HFI. This function seeks to maximize performance (e.g., reducing cost or time) but only when the probability of remaining within the feasible region exceeds a user-defined threshold, $\pi$.
By switching between these two modes, the framework avoids the common pitfall of "greedy" optimization, where an algorithm pursues a high-reward point that has a high likelihood of failure.
2.2 Status-Aware and Batch Optimization
To account for the realities of industrial hardware, the framework includes a Status-Aware calibration step. This allows the model to incorporate measurable machine data—such as gun voltage in plasma spraying—to account for equipment drift and wear over time. Furthermore, the implementation of Batch Bayesian Optimization allows the system to suggest multiple experimental points simultaneously, facilitating parallel testing and further reducing the temporal cost of configuration.
3. Experimental Evaluation
The framework was rigorously tested across three distinct domains:
- Synthetic Benchmarks: The algorithm was compared against the state-of-the-art "Expected Constrained Improvement" ($EI_C$) on three standard 2D constrained optimization problems.
- Atmospheric Plasma Spraying (APS): In both simulated and real-world settings, the goal was to regulate coating microhardness and porosity while minimizing a "stress index" related to machine wear.
- Fused Deposition Modeling (FDM): Real-world 3D printing experiments focused on minimizing print time for liquid-crystal polymers while maintaining a surface roughness constraint of less than 10 $\mu$m.
4. Key Insights and Results
The experimental results highlight a "Safety First" paradox: by being more cautious initially, the algorithm reaches the global optimum faster.
4.1 Efficiency and Feasibility
In APS simulations, the proposed method reached the optimal configuration using 43% fewer evaluations than standard $EI_C$. More importantly for manufacturing safety, the framework produced 2.25 times more feasible samples during the search. This reduction in "junk" samples translates directly to saved material and reduced machine downtime.
4.2 The Role of the Risk Threshold ($\pi$)
The parameter $\pi$ serves as a "risk dial" for engineers. A higher $\pi$ value (e.g., 0.8) forces the algorithm to be highly conservative, while a lower value (e.g., 0.3) allows for more aggressive exploration. The study found that values within this range provide a robust balance for most industrial applications.
5. Conclusion and Future Work
This research demonstrates that sample efficiency in manufacturing optimization is best achieved through a feasibility-aware approach. By accounting for machine status and prioritizing the discovery of feasible regions, the proposed BO framework offers a practical path toward autonomous process configuration.
While the framework is robust, future work will focus on refining the selection of Gaussian Process kernels and extending the switching logic to multi-objective optimization scenarios, where multiple competing quality metrics must be optimized simultaneously.
Summary updated