Session: 06-10-01: Advanced Manufacturing in Aerospace Engineering

Paper Number: 173247

Optimization of Fused Filament Fabrication (Fff) Process Parameters via Digital Twin: A Non-Invasive Framework 

Fused Filament Fabrication (FFF) has become a widely adopted Additive Manufacturing (AM) technique due to its accessibility and versatility. However, FFF suffers from process-induced defects that compromise geometric accuracy and mechanical performance. Common as-built defects include dimensional inaccuracies, layer shifting, surface roughness, voids, under-extrusion, and warping. These deviations, especially at the meso-scale, hinder the reliability and reproducibility of as-built components. Real-time, in-situ monitoring of the printed geometry remains one of the major technical barriers to optimizing process parameters and minimizing such defects.

This research presents a non-invasive digital twin framework to experimentally track, evaluate, and optimize process-induced geometric errors in FFF. The proposed method targets the Bambu Lab X1C printer and integrates computer vision, hardware augmentation, and software automation to achieve synchronized tracking between the virtual model (sliced G-code geometry) and the physical build. A red laser is mounted near the nozzle with a fixed offset and serves as a visible tracer during active extrusion. The printer’s real-time info panel, which displays the current layer number out of the total, is parsed frame-by-frame from a top-down camera feed. This enables accurate synchronization between the video footage and the sliced print layers.

The exposed laser trajectory is extracted and stored in (X, Y) coordinate space at each Z-height, forming a temporal point cloud of the printed path. This will be compared directly to the layer-by-layer geometry extracted from the corresponding G-Code file embedded in the sliced geometry file. The deviation between the recorded and true geometry is quantified at each layer and aggregated across the full print height, forming a geometric error map. This dataset becomes the foundation for future optimization of process parameters, such as extrusion multiplier, layer height, nozzle temperature, bed temperature, print speed, and acceleration.

The system is built using Python and integrates both automation and AI-assisted scripting components. STL slicing and printer communication, executing component slicing and launch import commands via Bambu Connect protocols, respectively. These scripts bypass GUI interaction and automate the G-code pipeline. In parallel, hands-free voice and text command parsing via OpenAI’s API and MediaPipe-based gesture control. While these AI layers are not essential to the core tracking and error mapping functions, they demonstrate extensibility for future deep learning control systems and optimization loops (e.g., inverse design, topology optimization).

Preliminary results from the laser exposure tracking show strong promise in capturing per-layer paths with sufficient fidelity for geometric reconstruction. The layered coordinates, stored in CSV format, enable immediate alignment with sliced model data for error quantification. The results confirm the feasibility of closed-loop geometry monitoring using simple hardware additions and software-driven calibration without interfering with the native printer firmware or build environment.

Holistically, this work provides a robust and scalable foundation for real-time defect feedback and process parameter optimization in FFF and is extensible to other AM mechanisms. The framework is flexible and adaptable for future integration with optimization algorithms and Machine Learning (ML) models to further enhance part quality, print reliability, and autonomous manufacturing capabilities.

Presenting Author: Pinar Acar Virginia Tech

Presenting Author Biography: Pinar Acar is an Associate Professor at the Mechanical Engineering Department of Virginia Tech. Her research interests include materials design, multi-scale materials modeling, design optimization, uncertainty quantification, and machine learning.

Authors:

Pinar Acar Virginia Tech
Carlos Vera Virginia Tech
Mohamed Elleithy Virginia Tech