Source-linked AI summary
Physics-Informed and Hybrid Machine Learning in Additive Manufacturing: Application to Fused Filament Fabrication
Berkcan Kapusuzoglu, Sankaran Mahadevan
TL;DR
The paper addresses limitations of standalone physics-based and ML models for predicting FFF part quality. It develops physics-informed and hybrid DNN strategies that combine physical constraints, multi-physics model outputs, and experimental data. The numerical results show improved accuracy and physically meaningful predictions, including accurate predictions with smaller experimental datasets.
Problem
Physics-based models can be computationally demanding or incomplete, whereas ML models can violate physical laws; the paper addresses this gap for FFF quality prediction.
Method
The paper develops eight combinations of three strategies: physics-informed loss functions, physics-model outputs as DNN inputs, and physics-model pre-training followed by experimental-data updating.
Results
The numerical results show improved prediction accuracy, physically meaningful results, and accurate model predictions with smaller amounts of experimental data.
Takeaways & Limitations
The approach combines ML’s ability to learn complex process-material-geometry relationships with physics knowledge while correcting approximations in the physics-based model.
Takeaways & Limitations
The study’s generalization to different part geometries, printers, and materials remains future work, and the relative weighting of experimental and physics-model data requires further investigation.
Abstract
from arXiv · showhide
This article investigates several physics-informed and hybrid machine learning strategies that incorporate physics knowledge in experimental data-driven deep-learning models for predicting the bond quality and porosity of fused filament fabrication (FFF) parts. Three types of strategies are explored to incorporate physics constraints and multi-physics FFF simulation results into a deep neural network (DNN), thus ensuring consistency with physical laws: (1) incorporate physics constraints within the loss function of the DNN, (2) use physics model outputs as additional inputs to the DNN model, and (3) pre-train a DNN model with physics model input-output and then update it with experimental data. These strategies help to enforce a physically consistent relationship between bond quality and tensile strength, thus making porosity predictions physically meaningful. Eight different combinations of the above strategies are investigated. The results show how the combination of multiple strategies produces accurate machine learning models even with limited experimental data.
1 Introduction
AM process optimization is costly and time-consuming, while physics-based and ML models each have important limitations. This paper develops physics-informed and hybrid DNN strategies for predicting FFF bond quality and porosity using physics knowledge, experimental data, and multiple model combinations.
- Motivation: Trial-and-error AM optimization requires repeated experiments for each new design, making it expensive and time-consuming.Model-based methods have therefore gained attention for optimizing process parameters.
- Motivation: Physics-based AM models require substantial computational effort and may need calibration for each process because they use approximations and incomplete representations of complex physics.Different models may also be needed for separate manufacturing sub-stages or phenomena.
- Motivation: ML models can represent complex AM relationships with sufficient data but may produce predictions inconsistent with physical laws when physical constraints are omitted.The paper motivates combining physics-based and ML approaches to improve accuracy and physical consistency.
- Approach: The paper investigates three strategies: physics-informed loss functions, physics-model outputs as DNN inputs, and physics-model pre-training followed by experimental-data updating.These strategies are evaluated in eight combinations for FFF porosity and bond-quality prediction.
- Contribution: The strategies target two physical quality metrics, neck diameter and porosity, while using multiple physics-based loss functions to improve data-driven AM models.This extends the physical quantities addressed compared with previous studies.
- Approach: The proposed models use physics constraints, physics-based models, and experimental data to predict porosity in FFF parts.An enhanced physics-based model also accounts for realistic filament geometry and geometry changes during printing.
2 Methodology
The methodology combines coupled thermal and polymer-sintering physics models, experiments, and PIML models to predict FFF bond formation and porosity. It integrates physical constraints, physics-model outputs, and experimental data through three strategies and their combinations.
- The methodology has three components: physics-based models, experiments, and construction of PIML models for predicting FFF bond formation and mesostructure.
- Physics-based models: The thermal model predicts filament temperature evolution, which feeds the sintering model to estimate porosity and bond quality.The thermal calculation considers material properties, part geometry, and process parameters.
- Bond formation modeling: The enhanced sintering model represents realistic filament geometry and allows filament geometry to change during printing.The model uses a filament cross-section consisting of a rectangle and two half circles, with geometry evolving during sintering.
- Bond formation modeling: The enhanced methodology produced smaller error than Gurrala et al.’s model but still relies on assumptions because sintering is not the only bond-formation phenomenon.This limitation motivates using deep learning to further enhance prediction accuracy.
- Physics-informed loss functions: Physics-informed losses combine the regular DNN training loss with weighted physics-based losses to enforce physically consistent predictions.Five physical relationships constrain predictions of dimensionless neck diameter and porosity.
- PIML strategies: PIML integrates physics through loss-function constraints, physics outputs as DNN inputs, or physics-model pre-training followed by experimental-data updating.The pre-training strategy uses a larger physics-generated input set than the physics-output-as-input strategy, but its advantage is limited when physics simulations are computationally expensive.
- PIML strategies: The investigated models include combinations of the three strategies, with Models 5, 6, and 7 combining pairs and Model 8 combining all three.
3 Implementation of PIML to FFF
The study implements and evaluates eight PIML models for FFF using experiments, physics simulations, and neural-network strategies. Physics-informed models improve physical consistency and accuracy, especially with limited experimental data.
- Experimental and model setup: Eight PIML models are evaluated using FFF experiments, physics-based simulations, and varying amounts of training data.Models are trained with 4, 6, 8, 10, or 20 experimental data points and tested on data from 19 parts.
- Experimental and model setup: The experiments manufacture unidirectionally aligned ABS parts across extrusion temperatures of 210°C–260°C and speeds of 15–46 mm/s.Latin hypercube sampling generates 20 process-parameter combinations.
- Model performance: Physics-based loss functions produce lower average RMSE and more physically meaningful results than models using physics-model outputs alone.The physics-output-only hybrid model performs similarly to the standard DNN, whereas loss-function strategies improve generalization.
- Model performance: Pre-training improves performance more strongly when fewer observed data are available and provides a more physically consistent initialization.This benefit occurs for updated models even without explicit physics constraints.
- Model performance: Models using physics-based loss functions produce physically consistent predictions and generalize to configurations unseen during training.The reported generalization benefit applies to models 2 and 4–8.
- Model performance: The combined strategy produces monotonic relationships between predicted porosity and neck diameter, unlike several models with physical inconsistencies.The test set contains 19 FFF parts; model 1 includes a negative porosity prediction of −0.025.
4 Conclusion
The paper develops physics-informed and hybrid models for FFF bond quality and porosity prediction by combining physics-based models, experiments, and DNN strategies. Results indicate improved accuracy and physical meaningfulness with smaller experimental datasets, while broader validation remains future work.
- Method: The study combines physics-based loss functions, physics-model outputs as DNN inputs, and physics-based pre-training updated with experimental data.The loss functions exploit the relationship between FFF bond quality and tensile strength.
- Method: The proposed sintering model improves on earlier models by representing realistic filament geometry and geometry changes during printing.It achieves a smaller error than Gurrala et al.’s model but still assumes sintering is the only relevant bonding phenomenon.
- Results: Physics knowledge improves prediction accuracy and physical meaningfulness while enabling accurate predictions with smaller amounts of experimental data.The approach also uses machine learning to represent complex process–material–geometry relationships and correct physics-model approximations.
- Scope and future work: Future evaluations should use higher-dimensional inputs, different geometries, other printers and materials, and investigate weighting between experimental and physics-model data.These extensions are identified as future work for assessing generalization and differing source credibility.
Conflict of Interest
The authors report no conflict of interest.
- The corresponding author states that there is no conflict of interest.