Page 43 - IJAMD-2-2

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

3.1.2. Property optimization combining six different ML algorithms to quantitatively

Beyond process optimization, ML has proven valuable forecast surface roughness. This approach underscores how
for predicting or enhancing the mechanical and physical sensor-based ML models can enhance predictive accuracy
properties of 3D-printed parts. One study applied ML to beyond conventional parameter-driven methods.
optimize the mechanical properties of FDM-fabricated ML has been instrumental in optimizing the properties
components, focusing on layer height, printing speed, of functional materials such as thermoelectric composites.
103
and printing temperature. By comparing RF, SVM, and Recent research has employed ML-assisted 3D printing
k-nearest neighbors (K-NN) models for tensile strength to fabricate thermoelectric materials with ultrahigh
prediction, the authors found that K-NN achieved the performance at room temperature, enhancing energy
highest accuracy. They then used a sequential least squares conversion efficiency through optimized microstructural
programming-based optimization algorithm to determine control (Figure 4B). In vat photopolymerization, ML has
109
optimal printing parameters. This ML-based approach also been used to optimize functional polymer composites,
enables the optimization of mechanical performance while including mechanoluminescent materials. Studies have
reducing reliance on purely experimental approaches. shown that ML-based optimization in SLA and DLP
A related study developed an ML model to predict the printing can enhance light-emitting properties by refining
hardness of additively manufactured ABS, demonstrating resin formulations and curing parameters, contributing to
the feasibility of using ML for estimating physical advancements in smart materials for sensing and display
properties beyond tensile strength. 104 applications. Moreover, high-throughput ML approaches
110
A related investigation leveraged ML to predict the have been integrated into the design of photodegradable
mechanical strength of biomedical structures printed hydrogels, enabling rapid formulation screening and
through FDM. In this case, polydopamine (PDM)- optimization for biomedical applications (Figure 4C). 111
105
coated PLA plates were tested for mechanical performance 3.1.3. Design optimization
using RF, K-NN, AdaBoost, Decision Tree, and LSTM
models. Input variables included infill density, PDM Functional polymers – often referred to as stimuli-
coating parameters (submersion time, shaker speed), and responsive, active, smart, or intelligent polymers – can
coating solution concentration. Similarly, hierarchical ML undergo shape change in response to external stimuli such
techniques have been explored to enhance the fidelity of as heat, electric or magnetic fields, or pH. 57,65,72,112-114 Owing
biopolymer-based 3D prints, addressing challenges in to advances in sophisticated manufacturing methods, these
achieving consistent mechanical and structural integrity polymers can now be designed with increasing complexity.
in biofabrication (Figure 4A). Another study combined Nevertheless, achieving specific shape transformations or
106
ML with sensor data to predict the mechanical properties functionalities often relies on labor-intensive, trial-and-
of FDM products. Employing an LSTM-based deep error procedures, incurring substantial time and expense.
107
learning model, the authors accounted for thermal and To address these challenges, recent work employed an
mechanical fluctuations at the layer level. Real-time data RNN to predict deformation patterns from a given material
– such as infrared readings, temperature measurements, distribution. Specifically, a bi-layered active composite
115
signals from thermocouples, and accelerometer outputs beam composed of two materials with differing thermal
– were fed into the LSTM network to analyze how nozzle expansion coefficients was optimized by integrating ML
temperature, print speed, and layer height influence tensile and EA. Conventional finite element-based optimization
strength. Compared with traditional ML models (RF, is often prohibitively expensive for complex geometries,
SVR), the LSTM approach demonstrated up to a 24.3% whereas the RNN-based approach significantly reduces
improvement in the coefficient of determination (R²). computational costs. By predicting the deformation
Furthermore, layer-wise relevance propagation revealed behavior of a bilayer active composite beam, the system
that layer height was the single most influential factor on could leverage EA to propose optimal material layouts
tensile strength. that govern shape change via volumetric expansion
Similarly, ML and sensor data have been applied to mismatch. This framework achieved over 99% reduction
predict surface quality in FDM. Traditional process in computational expenses compared with typical finite
108
parameters (e.g., printing speed, layer height) alone proved element workflows without sacrificing target-shape
inadequate for capturing variability in surface roughness. To fidelity. Moreover, by coupling this pipeline with computer
address this limitation, a combination of infrared sensors, vision techniques, the authors demonstrated the ability to
thermocouples, and accelerometers was employed to collect convert hand-drawn sketches into automatically generated
real-time data, followed by an ensemble learning model 4D-printed structures.

Volume 2 Issue 2 (2025) 37 doi: 10.36922/IJAMD025130010