Page 12 - IJAMD-2-2

International Journal of AI for
Materials and Design AI-driven material development for AM

random forest (RF) algorithm to predict the fatigue life human intervention. Lee et al. used the GAN model to
40
of a steel sample by analyzing a fatigue test dataset of 439 generate additional data to solve the problem of insufficient
steels containing nine alloying elements, and from the samples, which successfully improved the accuracy of the
437 possible combinations to select the composition and phase prediction of high-entropy alloys and revealed the
treatment of steel with the best fatigue life. In addition, key design parameters, providing a new insight for the
Navarrete et al. used ML models to predict the static yield design and discovery of high-entropy alloys.
35
stress of mixed cement pastes containing supplementary
cementitious materials by collecting datasets from previous 3.3. Other AI techniques
experimental work, and have compared the prediction Large language models (LLMs) are a class of AI models
accuracies of different ML models, including multilayer extensively applied in natural language processing. LLMs
perceptron, RF, and support vector regression. capture complex linguistic patterns and structures, enabling
These studies, although not related to AM, have multilingual processing and contextual understanding.
demonstrated the capabilities of ML in material property Their capabilities extend beyond language applications,
prediction and process optimization. By extracting with emerging studies demonstrating their potential in
41
complex relationships from extensive experimental materials science and engineering. For example, Buehler
datasets, ML models improve decision-making, process developed MechGPT, an LLM-based framework designed
efficiency, and reliability. As ML techniques continue to to simulate and analyze mechanical behavior and failure
evolve, their integration with computational modeling and mechanisms, improving the predictive accuracy of material
experimental validation will further accelerate the design properties and supporting the design of novel materials.
of next-generation materials and enhance manufacturing Genetic algorithms (GA) provide an effective strategy
intelligence. for global optimization, employing selection, mutation,
and crossover to explore complex solution spaces. In the
3.2. Deep learning (DL) study of Shen and Buehler, they used GA to further
39
DL leverages multilayer neural networks to extract features, optimize the microstructure of the model-designed
identify patterns, and model complex relationships materials to improve their mechanical properties. Nazar
in high-dimensional data. Compared to traditional et al., on the other hand, combined GA with the gene
42
36
ML methods, DL excels in automatic feature learning, expression concept to develop a new gene expression
making it particularly effective for image-based analysis, programming model to predict the plastic viscosity and
sequential data modeling, and generative design. Common yield stress of fresh concrete used for the MEX process
DL architectures include convolutional neural networks by analyzing six key factors, such as cement, sand, water,
(CNNs), recurrent neural networks, long short-term different sizes of coarse aggregate, and superplasticizers,
memory, generative adversarial networks (GANs), and to improve the accuracy and efficiency of predicting the
artificial neural networks (ANNs). rheological parameters of concrete.
DL extends data-driven approaches in material 4. Data collection and pre-processing
development, playing a key role in microstructure-
property correlation, process design, and automated A typical workflow of AI-driven material development in
material discovery. For example, Gu et al. used DL AM is illustrated in Figure 5. The workflow begins with
37
for the design of layered materials and trained a finite the collection of raw data from various sources, such as
element analysis (FEA) model to obtain high-performance simulations, experiments, literature, and databases. The
materials. Specifically, CNN was used to predict the raw data undergoes pre-processing to refine it for AI
mechanical properties of composites and combined it applications. Subsequently, AI algorithms are applied to
with a self-learning algorithm to optimize the design of train models that predict material behavior or suggest
the materials. Li et al. investigated the joint effects of the optimal material compositions. The workflow iteratively
38
size, depth, distribution, and orientation of defects on the integrates feedback from experiments and simulations to
fatigue life of AM-built Ti-6Al-4V alloys by constructing refine the dataset and improve prediction accuracy.
an ANN-based model, revealing how these factors interact Data collection and pre-processing are fundamental
with each other, which in turn affects the durability of the to the successful implementation of this workflow. High-
Ti-6Al-4V alloy. Shen and Buehler developed a novel quality and comprehensive datasets enable AI models to
39
unsupervised GANs (specifically StyleGAN) approach, identify underlying correlations and outliers that would be
trained with input unlabeled data, to construct a latent difficult to discern through traditional analysis. However,
space that is free to be explored for material design without the heterogeneity and complexity of data sources –

Volume 2 Issue 2 (2025) 6 doi: 10.36922/IJAMD025100007