Interactive 3D visualization of surface roughness predictions in additive manufacturing: A data-driven framework

Imagine you are baking a very complex, multi-layered cake. You want the outside to be perfectly smooth, but because you are stacking layers of frosting, the sides end up looking a bit like a staircase. In the world of 3D printing (specifically Fused Filament Fabrication, or FFF), this "staircase effect" is called surface roughness.

The problem is that this roughness isn't the same everywhere. A flat top might be smooth, but a steep side might be bumpy. And if you change your oven temperature or how fast you pipe the frosting, the texture changes again. Usually, figuring out the perfect settings requires a lot of trial and error—printing the part, measuring it, realizing it's too rough, changing the settings, and printing again. This wastes time, money, and plastic.

This paper introduces a smart, interactive crystal ball that predicts exactly how rough your 3D print will be before you even start printing.

Here is how they built it, explained simply:

1. The "Taste Test" (The Experiment)

To teach a computer how to predict roughness, the researchers needed a lot of data. They couldn't just guess; they had to print things and measure them.

The Recipe: They designed a special test object that looked like a staircase with 18 different steps, ranging from flat (0°) to almost upside down (170°).
The Variables: They changed 7 different "ingredients" in their recipe, like how thick the layers were, how hot the plastic was, and how fast the printer moved.
The Result: They printed 87 different versions of this object and measured the roughness of every single step. This gave them a massive dataset of 1,566 measurements. Think of this as a giant cookbook of "If I do X, the result is Y."

2. The "Brain" (Machine Learning)

They fed all this data into a type of computer brain called a Multilayer Perceptron (MLP).

How it works: Imagine a child learning to recognize animals. You show them a picture of a cat and say "cat," then a dog and say "dog." Eventually, the child can guess what a new animal is.
The AI's Job: This computer brain learned the complex, non-linear relationship between the "ingredients" (printing settings) and the "staircase angle" (geometry) to predict the final smoothness.

3. The "Imagination Engine" (Data Augmentation)

There was a problem: 87 prints is a lot of work, but for a super-smart AI, it's actually a tiny library. The AI might get confused if it hasn't seen a specific combination of settings before.

The Solution: They used a clever trick called a Conditional Generative Adversarial Network (CGAN).
The Analogy: Imagine you have a small collection of photos of cats. You want to show your friend more examples, but you don't have a camera. Instead of just copying the photos (which is boring), you ask a "fake artist" to draw new photos of cats that look real but are slightly different. A "detective" checks them to make sure they aren't obvious fakes.
The Result: The AI generated thousands of fake but realistic data points based on the real ones. This filled in the gaps in their "cookbook," allowing the prediction model to become much smarter and more accurate without needing to print more physical objects.

4. The "Magic Mirror" (The Interactive Tool)

The best part isn't just the math; it's the tool they built for regular people to use.

How it works: You go to a website and upload your 3D model (like a toy car or a vase). You tell the system, "I want to print at 200°C with 0.2mm layers."
The Visualization: The system instantly paints your 3D model with a color map.
- Blue/Green areas: "This part will be smooth."
- Red/Orange areas: "Warning! This part will be bumpy."
The Power: You can rotate the model in 3D space. As you turn it, the colors change instantly! You can see, "Oh, if I print it standing up, the side is rough, but if I lay it on its side, it becomes smooth."

Why This Matters

Before this, printing a complex part was like driving blindfolded. You'd print it, hope for the best, and if it was ugly, you'd start over.

This framework is like giving the driver night vision goggles. It lets you:

See the future: Know exactly where the rough spots will be before you spend a single gram of plastic.
Experiment instantly: Try different angles and settings in seconds on a screen instead of hours in a printer.
Save resources: Stop wasting materials on prints you know will fail.

In short, they built a system that turns the guesswork out of 3D printing, using a mix of real experiments, a little bit of AI imagination, and a colorful, interactive map to help anyone make smoother, better parts.

Here is a detailed technical summary of the paper "Interactive 3D visualization of surface roughness predictions in additive manufacturing: A data-driven framework."

1. Problem Statement

In Material Extrusion Additive Manufacturing (MEAM), specifically Fused Filament Fabrication (FFF), surface roughness is a critical quality attribute that varies significantly across a single part. Unlike traditional manufacturing, where surface finish is often uniform, FFF parts exhibit non-uniform roughness due to the "stair-stepping" effect caused by layer-wise deposition.

Challenges:
- Complex Interactions: Roughness depends on a non-linear interplay between process parameters (e.g., layer height, temperature, speed) and geometric factors (specifically local surface inclination relative to the build direction).
- Data Scarcity: Generating comprehensive experimental datasets is labor-intensive and costly. Existing datasets often lack sufficient coverage of the joint space defined by process parameters and surface angles, limiting the generalizability of data-driven models.
- Lack of Practical Tools: Current slicing software offers limited predictive feedback. Engineers rely on trial-and-error or experience-based adjustments, leading to wasted material, prolonged iteration cycles, and inconsistent quality, especially for parts with complex geometries and varying inclinations.

2. Methodology

The authors propose an end-to-end, data-driven framework comprising four main stages:

A. Structured Experimental Design

Specimen Design: A dedicated test specimen with 18 planar faces spanning inclination angles from 0° to 170° (in 10° increments) was designed to isolate the effect of surface angle.
Design of Experiments (DoE): A three-level Box–Behnken Design (BBD) was employed to explore the parameter space efficiently.
Parameters: Seven key process parameters were varied: Layer Height, Extrusion Temperature, Outer Wall Speed, Infill Density, Wall Thickness, Bed Temperature, and Fan Speed.
Data Collection: 87 specimens were printed (using PLA and a Bambu Lab A1 printer) without support structures. A contact stylus profilometer measured the arithmetic mean roughness ( $R_a$ ) on all faces, yielding 1,566 measurements.

B. Predictive Modeling (MLP Regressor)

Model Architecture: A Multilayer Perceptron (MLP) regressor was trained to map the 7 process parameters and the surface inclination angle to the predicted $R_a$ .
Training Protocol: The dataset was split into a hold-out test set (15%) and a training/validation pool (85%). Five-fold cross-validation with early stopping and hyperparameter optimization (via Optuna) was used to prevent overfitting.
Metrics: Performance was evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), Coefficient of Determination ( $R^2$ ), and Mean Absolute Percentage Error (MAPE).

C. Data Augmentation via CGAN

To address the limited size of the experimental dataset:

Technique: A Conditional Generative Adversarial Network (CGAN) was used to generate synthetic tabular data.
Conditioning: The generator was conditioned on the process parameters and surface inclination to ensure the synthetic $R_a$ values maintained the physical relationship with the inputs.
Integration: Synthetic samples were combined with real data (with reduced sample weights for synthetic data to prioritize empirical measurements) to train the MLP. The optimal augmentation ratio was determined via validation set performance.

D. Interactive Visualization Interface

Platform: A web-based Graphical User Interface (GUI) was developed.
Workflow: Users upload a 3D model (STL/OBJ), specify printing parameters, and adjust the build orientation.
Visualization: The system calculates the local surface inclination for every mesh facet, predicts $R_a$ using the trained MLP, and renders the results as an interactive colormap overlaid on the 3D geometry. This allows for real-time exploration of trade-offs between orientation and parameters.

3. Key Results

Predictive Performance

Real Data Only: The MLP trained solely on experimental data achieved an $R^2$ of 0.844 and an MAE of 2.200 $\mu$ m on the hold-out test set.
With CGAN Augmentation: Using a synthetic dataset three times the size of the original training set improved performance significantly:
- $R^2$ : Increased to 0.900.
- MAE: Reduced to 1.752 $\mu$ m.
- MAPE: Reduced to 10.89%.
Conclusion: Conditional augmentation effectively reduced prediction uncertainty, particularly in sparsely sampled regions of the parameter space, without introducing significant bias.

Feature Importance (SHAP Analysis)

Shapley Additive Explanations (SHAP) confirmed that Surface Angle and Layer Height are the dominant factors influencing roughness.
The analysis revealed that the relationship between surface angle and roughness is non-monotonic (complex), whereas layer height has a more consistent positive correlation with roughness. This validates that the model learned physically meaningful relationships rather than just statistical noise.

Visualization Utility

The web interface successfully demonstrated the ability to visualize spatial variations in roughness. Users could instantly identify "hotspots" of high roughness and iteratively adjust the build orientation to minimize them, demonstrating the tool's value for pre-print process planning.

4. Key Contributions

Comprehensive Dataset: Creation of a high-quality, reproducible dataset (1,566 measurements) covering a wide range of process parameters and surface inclinations, addressing a gap in existing literature.
Hybrid Data-Driven Framework: Successful integration of CGAN-based data augmentation with MLP regression to overcome data scarcity in manufacturing, proving that synthetic data can enhance generalization when validated rigorously.
Geometry-Aware Decision Support: Development of a practical, web-based tool that bridges the gap between abstract predictive models and engineering practice by visualizing roughness directly on 3D geometry.
Interpretability: Use of SHAP values to ensure the "black box" model aligns with known physical mechanisms of FFF (e.g., stair-stepping effects).

5. Significance and Impact

This research provides a robust solution to a persistent bottleneck in additive manufacturing: predicting and optimizing surface quality before fabrication.

Industrial Relevance: The framework reduces the reliance on costly trial-and-error, saving material and time. It enables "first-time-right" manufacturing for complex parts.
Accessibility: By deploying the model in a web-based GUI, the authors make advanced predictive capabilities accessible to practitioners who may not have deep expertise in machine learning or statistical modeling.
Scalability: The methodology is adaptable to other AM processes (e.g., SLA, SLS) and materials, provided the experimental design and metrology protocols are adjusted accordingly.

In summary, the paper presents a complete pipeline from experimental design to interactive application, demonstrating that data-driven approaches augmented by generative models can significantly improve the predictability and control of surface roughness in 3D printing.