Page 39 - IJAMD-2-3

International Journal of AI for
Materials and Design Biomimetic ML for AFSD aluminum properties

structures enables the integration of multifunctional employs customized neural network approximators to
materials, further enhancing thermal, mechanical, and model the coupled temperature evolution of the tool
corrosion resistance properties. The versatility of AFSD and the build during multilayer material deposition. The
also holds promise for the medical and energy sectors, model’s accuracy was validated through comparisons
where customized implants and components with tailored between predicted and actual temperature measurements.
properties could noticeably improve performance and Building on these advancements, the present study aims
service life.
to develop a novel biomimetic ML approach for predicting
Machine learning (ML) integration into the AFSD the mechanical properties of AFSD-fabricated aluminum
process is driven by the need to improve efficiency, alloy-walled structures. Specifically, the study focuses on
precision, and overall performance. This technology five aluminum alloys: AA2024, AA5083, AA5086, AA7075,
helps address key challenges in AFSD, such as managing and AA6061. The proposed method integrates genetic
numerous process parameters, anticipating material algorithm (GA)-optimized ML models with finite element
behavior, and ensuring product quality. AFSD involves a analysis to predict von Mises stress and logarithmic strain.
complex interplay of factors, including tool rotation speed,
feed rate, and axial force, all of which influence material AFSD holds significant promise as a manufacturing
flow, temperature distribution, and the mechanical technique for aluminum alloy structures. However, due
properties of the deposited layers. Traditional trial- to the complicated interaction of thermal and mechanical
and-error approaches to parameter optimization are processes inherent to AFSD, accurately predicting and
both time-consuming and costly. ML algorithms offer a controlling the mechanical properties of the final product
more efficient alternative by analyzing the vast datasets remains a significant challenge. Traditional experimental
generated during AFSD to identify optimal parameter approaches for optimizing process parameters and predicting
settings, thereby minimizing the need for extensive material behavior are both time-intensive and costly.
experimental runs. For example, Qiao et al. investigated Furthermore, the variability in material properties among
38
the use of ML approaches to optimize AFSD process these different aluminum alloys, namely AA2024, AA5083,
parameters for improved component design flexibility and AA5086, AA7075, and AA6061, further complicates this
performance. They used support vector machine, random task. Therefore, there is a pressing need for an effective and
forest (RF), and artificial neural network models to predict accurate method to predict critical mechanical parameters,
the mechanical properties – specifically microhardness such as von Mises stress and logarithmic strain, in AFSD-
and ultimate tensile strength – of AFSD-based AA6061 manufactured components across diverse aluminum alloys.
depositions. Key parameters such as temperature, force, This research seeks to address that need by introducing a
torque, rotation speed, traverse speed, feed rate, and layer novel biomimetic ML approach that integrates numerical
thickness were monitored using a self-developed process- modeling with GA-based optimization, offering a
aware kit. Among the models tested, the artificial neural potentially revolutionary solution for process optimization
network model demonstrated the highest accuracy, with and quality control in AFSD.
an R² of 0.9998, a mean absolute error (MAE) of 0.0050, 2. Materials and methods
and a root mean square error (RMSE) of 0.0063. The study
also identified feed rate and layer thickness as the most The simulations were conducted using Abaqus finite
influential factors on mechanical properties, contributing element software to model the AFSD process for five
24.8%/24.1% and 25.6%/26.6%, respectively. Zhu et al. aluminum alloys: AA2024, AA5083, AA5086, AA7075,
39
proposed a novel explainable artificial intelligence strategy and AA6061. The properties and typical applications of
that combines Bayesian learning with physics-based these alloys are summarized in Table 1.
surrogate models. They developed a physics-informed, Figure 1 illustrates the basic simulation setup. The
data-driven model capable of accurately forecasting modeling process begins with creating parts that represent
temperature distribution during AFSD by calibrating and both the substrate and the material to be deposited. Next,
updating these models using ML on in situ monitoring temperature-dependent material properties – including
data. The approach was validated through the AFSD of an
Al-Mg-Si alloy, resulting in rapid and accurate temperature density, specific heat, thermal conductivity, and elastic/
predictions with minimal reliance on physics-based plastic properties – are defined and assigned to both the
40
simulations and in situ measurements. Similarly, Shi et al. substrate and the deposited material.
introduced a physics-informed ML approach known as A step in Abaqus represents a specific phase or period
AFSD-Nets, which integrates heat generation and heat within the analysis during which specific loads, boundary
transfer effects to predict temperature profiles. AFSD-Nets conditions, and analysis procedures are applied. For the

Volume 2 Issue 3 (2025) 33 doi: 10.36922/ijamd.5014