MULTI-OBJECTIVE OPTIMIZATION UNDER UNCERTAINTY OF PART QUALITY IN FUSED FILAMENT FABRICATION

Additive manufacturing technologies like fused filament fabrication hold significant potential for producing complex parts directly from computer-aided design models. However, widespread industrial adoption remains limited by variability in final part quality. Traditional trial-and-error optimization through physical experimentation is expensive and time-consuming, while complex physics-based models are often computationally prohibitive for full-scale parts. This paper introduces a data-driven framework for multi-objective robust design optimization that explicitly accounts for uncertainty during the fused filament fabrication process.

The proposed approach proceeds in three main stages: experimental data collection, uncertainty-aware prediction modeling, and multi-objective robust optimization. First, laboratory experiments are conducted to print parts under varying process parameter configurations, after which quality metrics such as geometric inaccuracy and filament bond quality are measured offline. To model these quality characteristics, the authors construct Bayesian neural networks that use Monte Carlo dropout as an approximate variational inference technique.

Unlike deterministic neural networks, Bayesian neural networks estimate a posterior distribution over the network weights. This architecture allows the framework to decompose total prediction variance into two distinct components: epistemic uncertainty arising from limited training data and unknown model parameters, and aleatory uncertainty stemming from inherent observation noise and process variability. Stochastic forward passes yield predictive means and variances that feed directly into robustness-based design optimization formulations.

Using the statistical predictions from the trained Bayesian neural networks, the framework explores various robustness-based design optimization formulations. Key process parameters—including nozzle temperature, nozzle speed, and layer thickness—are optimized to simultaneously minimize geometric inaccuracy and maximize bond quality. Pareto surfaces are subsequently generated to evaluate the trade-offs between these competing objectives. The authors validate both the predictive accuracy of the Bayesian neural networks and the efficacy of the optimization methodology by physically manufacturing parts at the resulting optimal parameter settings.