Page 37 - IJAMD-2-2

International Journal of AI for
Materials and Design ML-driven optimization in additive manufacturing

Each ML paradigm presents specific strengths and deviations, and predicting potential defects. Material
32
trade-offs in terms of predictive accuracy, data efficiency, data, encompassing parameters such as composition,
and implementation complexity. Supervised learning viscosity, and thermal properties, influences printability,
generally achieves high accuracy when sufficient labeled microstructure evolution, and final part performance.
data are available, but it requires extensive data collection For instance, in vat photopolymerization, resin viscosity
and careful parameter tuning. In contrast, unsupervised and photoinitiator concentration affect curing dynamics,
learning offers greater data efficiency by leveraging whereas in PBF, powder morphology and packing density
unlabeled datasets. However, its outputs, such as clusters determine layer fusion and mechanical strength. 50,51 Quality
or anomaly scores, may lack the precision of supervised data, such as surface roughness, porosity, and mechanical
models due to the absence of ground-truth labels. Semi- properties, are typically obtained through non-destructive
supervised learning offers a compromise by combining evaluation techniques, including X-ray computed
limited labeled data with abundant unlabeled data to tomography (CT) and ultrasonic testing, providing ground
improve learning efficiency. For example, even a small truth for ML model validation. Geometric and design data
number of labeled defect images can significantly boost derived from computer-aided design (CAD) models and
detection accuracy when incorporated into a largely G-code instructions facilitate the analysis of layer-wise
unlabeled dataset. Reinforcement learning is particularly deviations, part distortions, and structural integrity. 23
suited to sequential decision-making tasks, such as real- Identifying which data types serve as input features
time process control, but its application in AM remains versus prediction targets is essential for designing effective
relatively limited due to the complexity of reward design ML workflows in AM. In most applications, process and
and the need for extensive experimentation. In practice, design data—such as machine sensor readings, processing
supervised CNN-based models have demonstrated over parameters, and CAD files—serve as input features. In
90% accuracy in classifying FDM defects when trained contrast, quality metrics (e.g., porosity and mechanical
with sufficient data, while unsupervised methods can still strength) typically serve as the output targets for prediction
effectively flag anomalies without any labels. Reinforcement or classification. Material data can play either role
learning, although promising for adaptive control, has thus depending on context, for example, known properties may
far been primarily explored in experimental setups. be used as inputs for performance prediction, or they may
In addition to purely data-driven methods, hybrid serve as outputs in formulation design tasks. In a defect
approaches that integrate physical modeling and ML– detection task, in situ, sensor images and environmental
often referred to as physics-informed or physics-based data constitute the input, while defect presence or severity,
ML–have gained attention for enhancing prediction verified by inspection, serves as the output. For property
reliability and reducing data requirements. For instance, prediction, inputs may include process conditions and
physics-informed neural networks have been applied in material composition, with outputs being target values
fused filament fabrication to incorporate thermal behavior such as tensile strength or elasticity. Proper alignment of
into property prediction models. In addition, physics- input–output structure helps guide algorithm selection
46
based simulations such as finite element analysis (FEA), and ensures that the ML approach is suited to the intended
computational fluid dynamics, and phase-field modeling application, be it predictive modeling, real-time anomaly
have been combined with ML to predict microstructure detection, or parameter optimization.
evolution, residual stress formation, and melt pool To extract meaningful insights from these datasets,
48
47
geometry in metal and ceramic AM processes. These various ML-driven data analysis techniques have been
49
approaches ensure physical consistency in ML predictions adopted. Image processing and computer vision models,
and offer scalable solutions for complex modeling processes including CNNs and You Only Look Once (YOLO)
in AM. networks, have been employed for real-time defect
The effectiveness of ML in AM largely depends on the detection in FDM by analyzing nozzle-near images. Time-
51
quality and availability of training data. Data generated by series analysis models, such as long short-term memory
AM can be categorized into process data, material data, (LSTM) networks, have been used to predict process drift
quality data, and geometric design data, each playing in PBF by analyzing laser scan data across multiple layers.
52
a crucial role in model development and validation. Hybrid approaches that combine ML with physics-based
Process data include sensor-based measurements such as simulations, such as FEA, have been developed to predict
temperature, pressure, humidity, and real-time monitoring residual stress formation and optimize support structures in
of melt pool dynamics in metal AM processes. These data metal AM. Furthermore, Bayesian optimization has been
53
points are critical for tracking process stability, detecting widely explored for process parameter tuning, reducing

Volume 2 Issue 2 (2025) 31 doi: 10.36922/IJAMD025130010