Page 16 - IJAMD-2-2

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

of unprocessed data can significantly compromise the in process parameters directly influence material structure
accuracy and reliability of AI models, necessitating robust and mechanical performance.
pre-processing methodologies to enhance data quality and In addition, AM datasets are often structured across
ensure compatibility with ML algorithms. A well-designed multiple spatial and temporal scales. Multiscale data
pre-processing pipeline ensures that datasets are not only integration is necessary to harmonize information collected
internally consistent but also scalable and representative at different levels, such as individual powder particles, melt
of the underlying material properties and behaviors. pools, printed layers, and entire components. This requires
Furthermore, it facilitates bias mitigation, minimizes the standardization of spatial resolutions and temporal
redundancy, and enhances predictive accuracy. sampling rates to enable meaningful feature extraction and
4.2.1. Fundamental data pre-processing techniques pattern recognition.
Data pre-processing typically begins with data cleaning, Feature extraction is a particularly critical aspect of
a process that involves handling missing values, detecting pre-processing in AM, as raw sensor data often requires
and correcting inconsistencies, and eliminating outliers. transformation before it can be effectively utilized in AI
In AI-based material research, missing values can arise models. For example, three-dimensional data derived
from X-ray, computed tomography (CT) scans, or
54
from incomplete experimental records or sensor failures microstructure imaging may be converted into structured
during process monitoring. Various imputation strategies representations through voxelization. Similarly, time-series
can be applied depending on data type and structure.
For numerical variables, mean or median imputation data from thermal sensors or acoustic emission monitoring
56
55
is commonly used, whereas categorical data are often may be analyzed using the Fourier or Wavelet transforms
to extract frequency-based features relevant to defect
addressed using the most frequent category or probabilistic detection and process stability. Dimensionality reduction
imputation techniques 53
techniques, such as principal component analysis, are
Feature encoding is another critical step, particularly frequently applied to condense high-dimensional feature
when dealing with categorical attributes, such as material sets while preserving essential information. 57
compositions, process parameters, and classification
labels. Conventional encoding techniques, such as one-hot 5. Applications of AI-enabled material
encoding, enable categorical variables to be represented development for AM
numerically without introducing unintended ordinal The development of materials with optimized properties
relationships. Feature scaling follows as a necessary step to is a central objective, driving advancements in structural,
ensure comparability across different feature dimensions. functional, and high-performance materials. The pursuit of
Common approaches include min-max normalization, improved mechanical strength, thermal stability, electrical
which scales features to a defined range, and standardization conductivity, and biocompatibility has led to extensive
(Z-score normalization), which transforms features to research into metals, polymers, and bioinks/biomaterial
have zero mean and unit variance, mitigating the effects inks.
of disproportionate feature magnitudes on model training.
This section explores the applications of AI in material
4.2.2. Pre-processing strategies for data in AM development for AM, with a particular focus on metals,

Unlike conventional structured datasets, datasets in AM polymers, and bioinks/biomaterial inks. For metals
are inherently multimodal, encompassing tabular datasets, and polymers, the discussion is structured into two key
high-resolution imaging, time-series process data, and aspects: Material design and performance optimization.
In the material design section, we examine how AI-driven
three-dimensional geometric representations. As a result, methods facilitate the discovery and optimization of novel
pre-processing strategies must be adapted to accommodate compositions/structures tailored for AM processes. The
these diverse data formats while ensuring cross-modality performance optimization section extends this analysis by
compatibility.
establishing a link between AM process parameters and
One of the key challenges in AM data pre-processing is material performance (Figure 6), leveraging AI to refine
data registration, which aligns information obtained from microstructural features and mechanical properties. For
different stages of the manufacturing process. For example, bioinks and biomaterial inks, the focus is on AI-assisted
ensuring that in situ monitoring data is correctly mapped formulation optimization, where balancing printability,
to corresponding microstructural characterization results rheological properties, biocompatibility, and mechanical
is essential for accurate process–property correlations. This integrity is crucial for ensuring successful bioprinting
alignment is particularly relevant in AM, where variations outcomes.

Volume 2 Issue 2 (2025) 10 doi: 10.36922/IJAMD025100007