💻 computer science

Hybrid Machine Learning and Physical Modeling of Feedstock Deformation During Robotic 3D Printing of Continuous Fiber Thermoplastic Composites

This paper presents a hybrid physics-based and data-driven modeling framework, combining Kelvin-Voigt viscoelasticity with stabilized neural ODEs, to accurately predict and validate continuous fiber thermoplastic feedstock deformation during robotic 3D printing, demonstrating robust generalization beyond training temperature ranges.

Original authors: Chady Ghnatios, Kazem Fayazbakhsh

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chady Ghnatios, Kazem Fayazbakhsh

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 trying to build a wall out of long, sticky strips of tape. You want to lay them down perfectly side-by-side to make a smooth, solid surface. But here's the problem: as soon as you lay the tape down, it starts to shrink, twist, and curl. It's like the tape is a living thing that changes its mind about its size the moment it hits the hot floor.

This is exactly the problem the researchers, Ghnatios and Fayazbakhsh, tackled. They are working on robotic 3D printing using special materials made of continuous carbon fibers (super-strong threads) wrapped in a plastic resin. When robots lay these fibers down, the material shrinks and warps, creating gaps or overlaps that ruin the final part.

Here is how they solved the mystery of the shrinking tape, explained simply:

1. The Mystery of the Shrinky Tape

The researchers noticed that when the robot lays down a strip of this material, it doesn't stay the same width. It shrinks from its original size (about 6.35 mm) down to about 5.3 mm. They wanted to know why and how much it shrinks so they could tell the robot exactly where to move to avoid gaps.

They identified three main "culprits" causing the tape to shrink:

The "Dry-Out": The tape holds a tiny bit of moisture that evaporates when heated, making it shrink.
The "Uncoiling": The tape was wound tightly on a spool before printing. When it gets hot, it releases that tension, like a spring uncoiling, which changes its shape.
The "Crystal Lock": The plastic inside the tape is like a liquid when hot. As it cools down, the molecules start arranging themselves into a rigid, crystalline structure (like water turning to ice). This process squeezes the material, making it shrink even more.

2. The Detective Work: The Lab Experiments

To understand these culprits, they put the tape in two special machines:

The "Thermometer" (DSC): This machine heated the tape up and measured how much heat it absorbed or released. It told them exactly when the plastic started turning into crystals (around 147°C).
The "Stretchy Tester" (DMA): This machine held a piece of tape and heated it up and cooled it down while measuring exactly how much it stretched or shrank. They did this twice: once to a "warm" temperature (120°C) and once to a "hot" temperature (180°C).

3. The Hybrid Detective: Physics + AI

The researchers didn't just guess; they built a "hybrid" detective system to predict the shrinkage. Think of it as a team of two experts working together:

Expert A (The Physics Professor): This part uses old-school math and physics rules (called a Kelvin-Voigt model) to understand how the material acts like a spring and a shock absorber. It knows how the tape reacts to heat and tension based on the laws of nature.
Expert B (The AI Student): This part uses a special type of Artificial Intelligence called a Neural ODE. Imagine a student who is really good at spotting patterns in a story. The AI looked at the data from the "warm" experiment (120°C) and learned the pattern of how the tape shrinks due to drying and uncoiling.

The Magic Trick: The researchers taught the AI student using only the "warm" data. Then, they asked the AI to predict what would happen in the "hot" experiment (180°C), a temperature it had never seen before.

The result? The AI got it right! It figured out that even though it hadn't seen 180°C, the pattern of shrinking was stable and predictable. It even correctly guessed that the shrinking would eventually stop (saturate) at higher temperatures. This proves the model is robust and can be trusted even when conditions change.

4. Putting It All Together: The Digital Twin

Finally, they combined the Physics Professor's rules with the AI Student's predictions to create a complete simulation. They fed this simulation the exact conditions of their robotic printer (how fast it moves, how hot the nozzle is, etc.).

The simulation acted like a "digital twin"—a virtual copy of the real printing process. It predicted exactly how much the tape would shrink in width and length. When they compared this prediction to a real tape they printed in the lab, the numbers matched up very closely.

The Bottom Line

The paper claims that by using this mix of physics and smart AI, they can now predict exactly how much a piece of 3D-printed tape will shrink before they even print it.

This is a big deal because, instead of guessing and running dozens of expensive tests to find the right settings, engineers can use this model to tell the robot exactly where to place the tape so that it fits perfectly, leaving no gaps and no overlaps. It turns a trial-and-error process into a precise, predictable one.

Technical Summary: Hybrid Machine Learning and Physical Modeling of Feedstock Deformation in Robotic 3D Printing

Problem Statement
Feedstock deformation during the 3D printing of continuous fiber-reinforced thermoplastic composites is a critical challenge that drives manufacturing defects such as gaps and overlaps. While thermal gradients and residual stresses are known contributors, the specific deformation of a single deposited tape—driven by residual stress relief, drying, and crystallization—remains complex to model. This single-path deformation directly impacts intrinsic process variables like fiber alignment and robotic deposition control. Existing approaches face limitations: purely data-driven models often lack reliability outside tested regions, while traditional physics-based models struggle with the coupled multiphysics of crystallization and microstructural property changes. The authors address the need for a robust framework that can predict tape dimensional changes to optimize path planning and minimize defects.

Methodology
The authors propose a hybrid modeling framework combining established physical models with data-driven approaches, specifically utilizing a stabilized Neural Ordinary Differential Equation (Neural ODE). The workflow proceeds through the following stages:

Experimental Characterization:
- Setup: A robotic 3D printing system (ABB IRB 1200) deposits continuous carbon fiber (CCF) prepreg tape (LM-PAEK matrix) onto a steel build platform.
- Material Testing: Differential Scanning Calorimetry (DSC) identified the glass transition temperature ( $T_g \approx 147^\circ\text{C}$ ) and crystallization peak ( $T_{pc} \approx 184^\circ\text{C}$ ). Dynamic Mechanical Analysis (DMA) was performed on tape specimens at two temperature regimes: $120^\circ\text{C}$ (below $T_g$ ) and $180^\circ\text{C}$ (above $T_g$ ). These tests measured dimensional changes (strain) during heating and cooling cycles.
Physical Modeling (Viscoelasticity):
- A Kelvin-Voigt viscoelastic model was employed to simulate the mechanical behavior of the prepreg.
- The model utilized orthotropic thermal and mechanical properties derived from manufacturer data and homogenization rules.
- Parameters (thermal expansion coefficients and viscous tensor components) were identified by fitting the simulation to the DMA-1 experimental data ( $120^\circ\text{C}$ ), excluding the initial heating phase where irreversible drying and residual stress release occur.
Hybrid Data-Driven Modeling:
- Initial Strain Release (Drying/Residual Stress): The discrepancy between the viscoelastic simulation and experimental data during the first heating phase of DMA-1 was attributed to drying and residual stress release. A stabilized Neural ODE was trained on this residual data to learn the time-dependent evolution of these initial strains.
- Crystallization-Induced Deformation: For the DMA-2 experiment ( $180^\circ\text{C}$ ), the difference between the total measured strain, the viscoelastic prediction, and the predicted initial strain release was defined as crystallization strain. A second stabilized Neural ODE was trained to model the evolution of these strains as the temperature exceeded $T_g$ .
Integration and Validation:
- The identified models (Kelvin-Voigt, initial strain release, and crystallization) were integrated into a holistic simulation framework using Proper Generalized Decomposition (PGD) for thermal and mechanical solving.
- This framework was applied to simulate a 280 mm tape deposition under robotic printing conditions ( $380^\circ\text{C}$ nozzle, $160^\circ\text{C}$ platform).
- The simulation results were validated against experimental measurements of a physically deposited tape.

Key Contributions

Hybrid Modeling Framework: The paper demonstrates a successful integration of physics-based viscoelasticity with stabilized Neural ODEs to capture complex, non-linear phenomena (drying, stress relief, crystallization) that are difficult to model purely analytically.
Extrapolation Capability: The study highlights the ability of the stabilized Neural ODE to generalize beyond its training region. Specifically, the initial strain release model, trained only on data up to $120^\circ\text{C}$ , successfully predicted behavior up to $180^\circ\text{C}$ , reaching saturation without retraining.
Decomposition of Deformation Mechanisms: The work isolates and quantifies the specific contributions of thermal expansion, initial strain release, and crystallization to the total feedstock deformation, providing a clearer understanding of the defect generation mechanisms in single-tape deposition.

Results

Model Accuracy: The hybrid model achieved high fidelity in predicting strain evolution. The relative error for the initial strain release prediction was 0.57% on the test set, and for crystallization strain, it was 2.43% on the test set.
Generalization: The model accurately reproduced the saturation of strain release at temperatures significantly higher than those used in training ( $180^\circ\text{C}$ vs. $120^\circ\text{C}$ training max).
Robotic Deposition Validation: When applied to the robotic printing scenario, the simulation predicted the lateral deformation of the tape with high accuracy. The simulated final width closely matched experimental caliper measurements, confirming that the model can account for the combined effects of thermal gradients, stress relief, and crystallization in a real-world deposition setting.

Significance
The authors claim this work constitutes a "cornerstone" in understanding feedstock deformation for robotic 3D printing of continuous fiber thermoplastics. The primary significance lies in the proposed "holistic approach" that enables the creation of a digital twin for the printing process. By accurately predicting the final dimensions of a deposited tape based on process parameters, the model allows for the optimization of the Tape Raster Offset (TRO) to prevent gaps and overlaps. This capability reduces the reliance on costly and time-consuming experimental trials to determine optimal printing parameters. Furthermore, the authors suggest that while the current study focuses on straight paths, the framework provides a foundation for future work on curvilinear paths and the manufacturing of complex structural parts.