Page 40 - IJAMD-2-2

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

Table 1. (Continued)
Motivation Materials AM method ML model Model inputs Remarks References
ABS FDM LR, DT, RF, AdaBoost Infill density, layer Predicting and optimizing 109
thickness, print hardness
orientation, raster
orientation
Polydopamine- FDM RF, K-NN, AdaBoost, Infill density, Predicting mechanical 110
coated PLA DT, LSTM submersion time, shaker strength (tensile and
speed, coating solution bending) of PDM-coated
concentration PLA structures
PLA FDM LSTM Infrared temperature, Predicting tensile strength of 112
thermocouple, PLA parts using process and
accelerometer, printing sensor data
parameters (extruder
temperature, printing
speed, layer height)
PLA FDM Ensemble learning (RF, Layer thickness, Predicting surface roughness 113
AdaBoost, CART, SVR, extruder temperature,
RR, RVFL) print speed, infrared
sensor, thermocouple,
accelerometer data
Acrylate DLP GPR, BO, AL Monomer composition Optimizing photopolymer 117
photopolymer resin ratios, Young’s modulus, resin formulations using
peak stress, ultimate active learning and Bayesian
strain, Shore A hardness optimization
Acrylate DLP Ensemble learning Resin composition Predicting multiple 118
photopolymer resin (K-NN, GPR, KR, RF, ratios, hardness, tensile mechanical properties of
MLP) strength, elongation at photopolymers
break
Design Acrylate DLP RNN+EA Material distribution Predicting shape change and 120
optimization photopolymer resin (binary voxelized optimizing material layout
structure) for 4D printing
Acrylate DLP ResNet (CNN-based) Voxel-level material Predicting shape 121
photopolymer resin distribution deformations and optimizing
material distribution
Abbreviations: ABS: Acrylonitrile butadiene styrene; AL: Active learning; ANN: Artificial neural network; BO: Bayesian optimization;
CART: Classification and regression tree; CGAN: Conditional generative adversarial network; CNN: Convolutional neural network; DANN: Domain
adversarial neural networks; DLP: Digital light processing; DT: Decision trees; EA: Evolutionary algorithm; EWMA: Exponentially weighted moving
average; FDM: Fused deposition modeling; GelMA: Gelatin methacrylate; GPR: Gaussian process regression; K-NN: k-nearest neighbor; KR: Kernel
ridge regression; LSTM: Long short-term memory; ML: Machine learning; MLP: Multi-layer perceptron; PLA: Polylactic acid; RF: Random forest; RNN:
Recurrent neural network; RR: Ridge regression; RVFL: Random vector functional link network; SLA: Stereolithography; SVM: Support vector machine;
SVR: Support vector regression UV: Ultraviolet; YOLO: You Only Look Once.

YOLOv3-Tiny model accelerated by the Open Neural model overfitting. Another work employed a dual approach
Network Exchange runtime achieved 70 FPS, underscoring combining imaging and sensor data. A CNN-based
75
the viability of high-speed in situ defect compensation. model classified the severity of interlayer delamination
In addition to image-based approaches, sensor signals from nozzle-near images, while signals from a strain gauge
have also been integrated into ML frameworks to predict attached to the build platform predicted warping that
defects. 75,83 In one study, environmental parameters might develop gradually (Figure 3A). The strain data were
(temperature, humidity, air pressure, and gas-particle interpreted using a physics-based model, which indicated
concentration) were monitored in real-time and used with that warping was likely to occur once the strain exceeded a
an XceptionTime model to classify FDM process states certain threshold.
automatically. Air pressure emerged as the most critical Compared with FDM, adopting ML for DIW is
83
feature, possibly because its relatively stable baseline made relatively new, but several studies have explored ML-based
small variations highly indicative and also helped prevent optimization of DIW processes. 84-86 A study has reported

Volume 2 Issue 2 (2025) 34 doi: 10.36922/IJAMD025130010