Physics-Informed Dynamical Modeling of Extrusion-Based 3D Printing Processes

This paper proposes a reduced-order, physics-informed dynamical model for extrusion-based 3D printing that balances high fidelity with computational efficiency, demonstrating strong agreement with CFD simulations while remaining suitable for real-time optimization and control.

The Gist

Imagine you are a master chef trying to pipe the perfect swirl of frosting onto a cake. You have a piping bag (the nozzle) and a moving conveyor belt (the cake stand). If you squeeze too hard, the frosting piles up; if you move the belt too fast, the swirl stretches out and breaks. To get it perfect every time, you need to know exactly how the frosting will behave in real-time.

This paper is about teaching a computer to be that master chef, but for 3D printing with cement (like building houses layer by layer).

Here is the breakdown of their work using simple analogies:

The Problem: The "Super-Computer" vs. The "Real-Time" Dilemma

Scientists already have "Super-Computers" (called CFD simulations) that can predict exactly how wet cement flows. They are like a high-definition, slow-motion movie of the frosting. They are incredibly accurate, but they take hours or days to run.

The Catch: You can't use a movie that takes 24 hours to render to control a robot arm that needs to move right now. If the robot waits for the computer to finish its calculation, the building will be ruined.

The Goal: The authors wanted to create a "shortcut" model. Think of it like turning that 4K slow-motion movie into a simple, fast sketch that a human can look at and instantly understand what's happening, without losing the important details.

The Solution: The "Three-Stage Relay Race"

Instead of trying to model the entire messy flow of cement at once, the authors broke the process down into three simple "stations" or sub-systems, like a relay race:

1. Station 1: The Nozzle (The Squeeze)

   • What happens: The cement is pushed through a tube.
   • The Analogy: Imagine water flowing through a garden hose. The speed depends entirely on how hard you squeeze the handle (the pump). The speed of the ground outside doesn't matter here.
   • The Model: They created a simple math rule: More squeeze = faster flow.

2. Station 2: The Gap (The Handoff)

   • What happens: The cement leaves the nozzle and hits the moving floor.
   • The Analogy: This is like a runner passing a baton to another runner who is already sprinting. The new runner (the cement) gets a little push from the old runner (the nozzle) but also gets dragged along by the sprinter (the moving floor).
   • The Model: This is the tricky part. It's a mix of the squeeze and the floor speed. They used a simple algebraic "bridge" to connect the two.

3. Station 3: The Layer (The Drag)

   • What happens: The cement sits on the moving floor.
   • The Analogy: Imagine a piece of tape being pulled across a table. Once it's on the table, how fast it moves depends almost entirely on how fast you are pulling the table, not how hard you pushed it onto the table.
   • The Model: The rule here is simple: The floor speed dictates the cement speed.

How They Taught the Model

They didn't just guess the math rules. They used a "Teacher-Student" approach:

• The Teacher: The slow, perfect "Super-Computer" (CFD) that ran 9 different scenarios (different squeeze strengths and different floor speeds).
• The Student: Their new, simple "Sketch" model.

They let the Student watch the Teacher solve the problems, then they tweaked the Student's math knobs (parameters) until the Student's answers matched the Teacher's answers almost perfectly.

The Results: Does the Sketch Work?

They tested the "Sketch" model in three ways:

1. The "In-Between" Test (Interpolation): They trained the model on "Hard Squeeze" and "Soft Squeeze," then asked it to predict "Medium Squeeze."

   • Result: It nailed it. It was like asking a student who studied the hardest and easiest math problems to solve a medium one. They got it right.

2. The "Outside the Box" Test (Extrapolation): They trained the model on "Soft" and "Medium" squeezes, then asked it to predict a "Super Hard" squeeze (something it had never seen).

   • Result: It struggled a bit. This makes sense! If you only practice driving at 20 mph, you might panic when you suddenly hit 80 mph. The model works best when predicting things within the range it has seen, or slightly below it.

3. The "Random Mix" Test: They threw random combinations of speed and squeeze at the model.

   • Result: It remained stable and accurate, proving it's robust enough for real-world use.

Why This Matters

This paper gives us a fast, lightweight engine for 3D printing cement.

• Before: You had to wait days to simulate a print, meaning you couldn't fix mistakes while printing.
• Now: With this model, a computer can predict what the cement will do in milliseconds. This allows for real-time control. If the printer sees the cement is about to sag, the computer can instantly adjust the speed or pressure to fix it, just like a chef adjusting their hand while piping frosting.

In short: They turned a slow, complex physics movie into a fast, simple sketch that a robot can use to build houses perfectly, one layer at a time.

Technical Summary

1. Problem Statement

Extrusion-based additive manufacturing (AM), specifically Direct Ink Writing (DIW) for cementitious structures, faces a critical trade-off between model fidelity and computational cost.

• High-Fidelity Models: Traditional approaches using Computational Fluid Dynamics (CFD) and Finite Element Methods (FEM) provide detailed insights into flow and stress but are computationally expensive. They are unsuitable for real-time monitoring, optimization, and closed-loop control.
• Data-Driven Gaps: Existing reduced-order models (ROMs) often lack physical grounding or fail to explicitly capture the evolution of the extruded strand as a function of governing process parameters.
• The Gap: There is a lack of computationally tractable, physics-based dynamical models specifically formulated for the flow dynamics of DIW processes that can support real-time control applications.

2. Methodology

The authors propose a Physics-Informed Reduced-Order Dynamical Model that bridges the gap between first-principles physics and data-driven efficiency. The methodology involves three main stages:

A. Physical Decomposition and Modeling

The printing process is decomposed into three cascaded sub-systems, each modeled with specific physics-based assumptions:

1. Sub-system 1 (Nozzle Interior):

   • Physics: Governed by Navier-Stokes equations for incompressible, Newtonian flow.
   • Derivation: The authors apply Conservative Averaging Method (CAM) to the momentum and continuity equations. By assuming a parabolic velocity profile and integrating over the cross-section, they derive a linear Ordinary Differential Equation (ODE) relating the inlet mass flow rate ($\dot{m}$) and pressure gradient to the average outlet velocity ($\bar{v}_1$).
   • Resulting Equation: $d\bar{v}_1/dt = \beta_1 p_{d1} + \beta_2 \bar{v}_1 + \beta_3 \dot{m}$.

2. Sub-system 2 (Nozzle-Substrate Gap):

   • Physics: Modeled as a transitional region where the extrudate interacts with the moving substrate.
   • Derivation: A simplified linear algebraic mapping is used to relate the nozzle outlet velocity to the inlet velocity of the next stage, capturing the entrainment effect of the moving plate.
   • Resulting Equation: $\bar{u}_2 = \beta_4 \bar{v}_1$.

3. Sub-system 3 (Deposited Layer):

   • Physics: Describes the flow on the moving build plate.
   • Derivation: Similar to Sub-system 1, CAM is applied to the momentum equation with boundary conditions defined by the moving plate velocity ($U_s$). This yields an ODE relating the input velocity and plate speed to the output velocity ($\bar{u}_3$).
   • Resulting Equation: $d\bar{u}_3/dt = \beta_5 \bar{u}_2 + \beta_6 \bar{u}_3 + \beta_7 U_s$.

B. Input-Dependent Parameterization

To address the non-linearities of the process (where dynamics change based on operating conditions), the model coefficients ($\beta$) are not treated as constants. Instead, they are parameterized as linear functions of the primary inputs:

• Coefficients for Sub-system 1 depend on the inlet mass flow rate ($\dot{m}$).
• Coefficients for Sub-system 3 depend on the plate velocity ($U_s$).
• This creates a structure: $\beta = \gamma \cdot \text{Input}$, where $\gamma$ are the tunable parameters.

C. Data Generation and System Identification

• Data Source: High-fidelity 3D CFD simulations were performed using ANSYS Fluent (multiphase flow, cement-based material vs. air) across nine distinct operating conditions (combinations of inlet velocities 50–70 mm/s and plate velocities 50–70 mm/s).
• Training: A nonlinear least-squares approach (using MATLAB) was employed to identify the parameters ($\gamma$ and $\beta_4$) by minimizing the error between the model's time-discretized output and the CFD ground truth data.

3. Key Contributions

1. Novel Reduced-Order Framework: Development of the first reduced-order model specifically for DIW flow dynamics that retains the structure of the Navier-Stokes equations while reducing complexity to ODEs and algebraic equations.
2. Physics-Informed Parameterization: The model does not rely purely on black-box fitting; it uses physics-derived structures where parameters are explicitly linked to physical inputs ($\dot{m}$ and $U_s$), enhancing interpretability and generalization.
3. Modular Architecture: The three-subsystem approach allows for modular analysis of the nozzle, gap, and deposition regions, facilitating targeted control strategies.
4. Real-Time Viability: The resulting model is computationally lightweight, making it suitable for integration into real-time control loops, unlike traditional CFD.

4. Results

The model was rigorously validated through interpolation and extrapolation scenarios:

• Interpolation (Mass Flow): When trained on extreme mass flow rates (low and high) and tested on intermediate rates, the model achieved high accuracy.
  • Sub-system 1 & 3: RMSE was very low, with error distributions tightly clustered around zero ($\pm 0.005$ m/s).
  • Sub-system 2: Showed slightly higher error due to its simplified algebraic nature but remained within acceptable bounds.
• Extrapolation (Mass Flow):
  • The model performed well when extrapolating to lower mass flow rates (trained on high flow).
  • Performance degraded when extrapolating to higher mass flow rates (trained on low flow), as high-flow regimes introduce stronger non-linearities and coupled effects not captured in the low-flow training data.
• Surface Velocity ($U_s$):
  • Low-end extrapolation: The model generalized exceptionally well to lower plate velocities (trained on high velocities), as low-velocity dynamics are effectively scaled-down versions of high-velocity dynamics.
  • High-end extrapolation: Accuracy decreased when predicting higher velocities, as the model struggled to capture the intensified shear and non-linear coupling present at extreme speeds.
• Randomized Partitioning: Under randomized train-test splits (varying from 67/33 to 33/67 ratios), the model demonstrated robust generalization, maintaining stable performance across heterogeneous operating conditions.

5. Significance

• Enabling Real-Time Control: This work provides a computationally efficient alternative to CFD, enabling the development of closed-loop control systems for 3D printing. This is crucial for maintaining geometric accuracy and print quality in real-time.
• Smart Manufacturing: The model serves as a foundation for "smart" manufacturing systems capable of adaptive process control, monitoring strand quality, and optimizing flow stability without the latency of high-fidelity simulations.
• Bridging Theory and Practice: By grounding the data-driven model in physical laws (Navier-Stokes), the authors ensure the model is not only accurate within the training domain but also physically plausible and interpretable, addressing the "black box" limitations of pure machine learning approaches.

In conclusion, the paper successfully demonstrates that a physics-informed, reduced-order dynamical model can accurately capture the transient flow behavior of extrusion-based 3D printing, offering a viable pathway for real-time optimization and control in civil infrastructure fabrication.