⚡ electrical engineering

Geometry-aware interpretable learning for predicting print time and filament consumption in fused deposition modeling

This study presents a geometry-aware interpretable machine learning framework that accurately predicts fused deposition modeling print time and filament consumption by leveraging a dataset of 1,500 simulated instances, achieving high predictive fidelity ( $R^2 > 0.94$ ) and revealing that mesh geometry and specific process settings are the primary determinants of these operational outcomes.

Original authors: Hoa-Cuc. Nguyen, Viet-Hung. Pham, Bich-Ngoc. Mach, Ngoc-Duong. Nguyen, Tran-Thi Hong, Thanh Q. Nguyen, Ma-Thi-Anh Thu

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hoa-Cuc. Nguyen, Viet-Hung. Pham, Bich-Ngoc. Mach, Ngoc-Duong. Nguyen, Tran-Thi Hong, Thanh Q. Nguyen, Ma-Thi-Anh Thu

Original paper licensed under CC BY 4.0 (https://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 a workshop where machines build three-dimensional objects layer by layer, melting plastic and laying it down with the precision of a pen drawing on paper. This is fused deposition modeling, a technology that has moved from specialized labs to school classrooms and small businesses because it is affordable and versatile. Yet, for all its accessibility, the process remains frustratingly unpredictable for many users. Before a machine starts working, an operator must guess how long the job will take and how much plastic it will consume. These guesses are often just educated hunches, leading to wasted time, interrupted schedules, or running out of material halfway through a project. The difficulty lies in the fact that the time and material needed are not determined by the machine's settings alone. They are the result of a complex conversation between the shape of the object being made and the instructions given to the printer. A small change in the design of the part can completely alter how the machine behaves, making simple rules of thumb unreliable.

Researchers at Thu Dau Mot University and several other institutions in Vietnam set out to solve this problem by teaching a computer to understand the relationship between the shape of an object and the machine's settings. They did not just look at the printer's dials and switches; they looked at the digital blueprint of the object itself. They gathered data from 1,500 simulated printing jobs, created from 150 different 3D shapes, each tested under ten different machine configurations. This allowed them to see how the same object would behave with different settings, and how different objects would behave with the same settings. They fed this information into a learning system designed to predict two specific outcomes: the total time the printer would run and the total length of plastic filament it would use. The goal was not just to make a prediction, but to understand which factors actually mattered most, so that users could make better decisions without needing to be experts.

The researchers found that the shape of the object is the single most important factor in determining both time and material use. The computer learned that the total surface area and the volume of the object are the primary drivers of the outcome. If an object is large or has a complex surface, it will take longer and use more plastic, regardless of how the machine is tuned. This finding challenges the common assumption that tweaking the machine's speed or temperature is the main way to control the process. While the machine settings do have an effect, their influence is secondary to the physical burden imposed by the object's geometry. The study showed that for predicting how much plastic is needed, the size of the object and the speed of the print head are the most critical factors. For predicting how long the job will take, the surface area of the object is the dominant factor, followed by the volume and specific settings like the number of outlines around the object and the thickness of each layer.

To make these predictions, the team built a system that combines two different types of learning algorithms. One algorithm, known as a random forest, acts like a committee of decision-makers, each looking at the data from a slightly different angle to find a stable answer. The other, called extreme gradient boosting, learns by correcting its own mistakes step by step, becoming more precise with every iteration. The researchers then combined the insights from both systems into a final prediction. This approach proved to be highly accurate. In their tests, the system predicted the print time and filament length with an error rate of less than ten percent for real-world cases. For example, when the system predicted a job would take 256 minutes, the actual time was 270 minutes. When it predicted 27 centimeters of plastic would be used, the actual amount was 25.9 centimeters. These small deviations are well within the range needed for practical planning, allowing a user to schedule a job or order materials with confidence.

Perhaps the most valuable part of the study was not just the accuracy of the numbers, but the clarity of the explanation. The system did not act as a black box that simply gave an answer; it revealed which variables were truly important. It showed that users do not need to obsess over every single setting on their printer. Instead, they can focus on a few key factors that the object's shape dictates, while leaving many other settings at their default values. This reduces the complexity of the task, making the technology more accessible to non-experts. The research suggests that the future of 3D printing support lies not in more complex tuning, but in understanding that the object itself is the main character in the story. By recognizing that the geometry of the part sets the stage, and the machine settings are merely the actors, users can make smarter decisions before they even press the start button. This approach transforms the process from a game of trial and error into a predictable, manageable workflow, bridging the gap between digital design and physical reality.

Technical Summary: Geometry-Aware Interpretable Learning for FDM Prediction

Problem Statement
Efficient Fused Deposition Modeling (FDM) is hindered by the nonlinear interaction between part geometry and process settings. Current operational practices often rely on heuristic tuning or trial-and-error, lacking a transparent framework to connect mesh geometry with controllable printer parameters. While prior research has focused on parameter optimization for specific materials or isolated post-print properties (e.g., mechanical strength), there is a gap in predicting pre-print operational outcomes—specifically print time and filament consumption—across diverse geometries. Existing predictive models often treat settings as the primary drivers, underrepresenting the structural burden imposed by the mesh, or they function as "black boxes" that lack the interpretability required for actionable configuration guidance.

Methodology
The study proposes a geometry-aware interpretable multi-output learning framework designed to jointly predict print time and filament consumption.

Dataset Construction: The empirical foundation consists of 1,500 printing instances derived from 150 unique 3D meshes. These instances were generated within a sustained, long-running printing environment operated since 2019, where print samples, parameter settings, and production records were continuously accumulated. The dataset structure involves evaluating each mesh under 10 distinct, expert-validated configuration settings. While the primary learning resource utilized simulation-supported records to ensure scale and diversity, the simulation parameters were verified against actual printing behavior, with a subset of over 3,000 available models physically printed to confirm consistency with real-world processes.
Input Variables (17 total):
- Geometry Descriptors (7): Area, Volume, Vertices, Faces, Height, Depth, and Width.
- Printer Settings (10): Top/Bottom/Outline/Skirt Layers, Skirt Outlines, Extrusion Width, Layer Height, Max Speed, Infill, and Internal Fill Pattern.
Data Conditioning: The pipeline includes categorical encoding, min-max normalization to handle scale differences between geometric and process variables, and outlier filtering using Isolation Forest to remove anomalous scaling behaviors.
Model Architecture:
- Base Learners: Random Forest (bagging) and Extreme Gradient Boosting (XGBoost, boosting) were trained independently.
- Fusion Strategy: A stacking-based multi-output regression approach was employed. The predictions from the base learners served as meta-features for a linear meta-learner, creating a fused model.
- Interpretability: Feature importance rankings were derived to distinguish between structural drivers (geometry) and controllable levers (settings).
Evaluation Protocol: The dataset was split 80:20 (train/test). Hyperparameters were optimized via Ray Tune and HyperOpt using 5-fold cross-validation. Performance was benchmarked against the individual base learners and a previously reported Convolutional Neural Network–Multilayer Perceptron (CNN-MLP) baseline. Metrics included RMSLE (for tuning), MSE, MAE, RMSE, and $R^2$ (for testing).

Key Results

Predictive Performance: The stacking framework achieved superior performance compared to both individual ensemble models and the CNN-MLP baseline.
- Overall Metrics: MSE = 0.0001929, MAE = 0.0063829, RMSE = 0.01361, $R^2$ = 0.9619.
- Target-Specific Performance:
  - Print Time: $R^2$ = 0.9645.
  - Filament Length: $R^2$ = 0.9449.
- The framework substantially outperformed the CNN-MLP reference, which reported an MSE of 0.0059 and MAE of 0.0350 for the combined targets.
Feature Importance & Interpretability:
- Geometry Dominance: Mesh geometry was identified as the primary driver for both outputs. For filament consumption, geometry variables accounted for ~72.6% of cumulative importance; for print time, this rose to ~79.9%.
- Key Drivers:
  - Filament: Volume and Area were the strongest predictors. Among settings, Speed was the most influential.
  - Print Time: Area and Volume were principal determinants. Among settings, Skirt Outlines, Layer Height, and Extrusion Width exerted the strongest effects.
Practical Validation: In representative real-print scenarios, the model demonstrated deviations within the sub-10% range (7% for filament, 5% for time), confirming its utility for pre-print planning. Sensitivity analysis confirmed that low-ranked variables (e.g., specific layer counts) had negligible impact on outcomes, validating the reduced-complexity configuration hypothesis.

Key Contributions

Geometry-Aware Framework: The study establishes a unified learning architecture that explicitly integrates mesh descriptors with process settings, treating geometry not as passive background but as an active determinant of operational burden.
Interpretability-Driven Configuration: By ranking feature importance, the framework identifies a reduced set of materially important configuration variables. This shifts the paradigm from optimizing the entire parameter space to focusing on a specific subset of high-leverage settings, simplifying decision-making for non-expert users.
Multi-Output Operational Intelligence: The work moves beyond single-output or post-print property prediction to jointly model pre-print scheduling (time) and material allocation (filament), providing a coherent basis for operational inference.

Significance and Claims
The paper claims that its primary significance lies not merely in achieving high predictive accuracy, but in transforming raw printing data into actionable, reduced-complexity printer-setting intelligence. By demonstrating that mesh geometry dominates the operational response, the study argues that efficient FDM planning requires a "geometry-first" perspective. The framework supports a transition from heuristic, trial-and-error setup to a transparent, data-driven decision process where users can rely on accurate pre-print estimates and simplified configuration guidance. The authors position this work as a step toward deployable support systems that bridge the gap between geometric design and process execution in additive manufacturing.

Technical Summary: Geometry-Aware Interpretable Learning for FDM Prediction

More like this