Page 13 - IJAMD-2-2

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

Figure 5. Workflow for artificial intelligence-driven material development in additive manufacturing

ranging from atomic-scale simulations to experimental Table 2. Common data types and sources in material
measurements – present significant challenges. These development
challenges include inconsistencies in data formats, Data source Data type Examples
missing and noisy values, and variability in measurement
techniques. Pre-processing mitigates these issues by Experiment Scalar Ultimate tensile strength, hardness,
relative density, and thermal
standardizing and transforming raw data, ensuring its conductivity
compatibility with AI algorithms. Time-series Stress-strain curve and thermography
4.1. Data collection Spectral X-ray diffraction, X-ray dispersive
spectroscopy, X-ray photoelectron
Data for material development are primarily derived spectroscopy, and Raman spectroscopy
from simulations and/or physical experiments. Overall, Image Scanning electron microscopy,
the types of datasets for material development can be transmission electron microscopy, and
broadly categorized into scalar, time-series, spectral, atomic force microscopy
image, categorical, and spatial data. Scalar data represents Categorical Phase and defect
single-value material properties, such as tensile strength Spatial Crystallographic texture, pore
and elastic modulus. Time-series data captures changes distribution, and filler dispersion
over time, such as stress-strain behavior during tensile Simulation Scalar Total energy (DFT), Gibbs free energy
testing. Spectral data reveals material composition and (CALPHAD), elastic modulus (DFT,
structure through techniques, such as X-ray diffraction. FEM), and stress/strain (FEM)
Image data mainly includes microstructure images, while Time-series Displacement evolution (FEM)
categorical data describes qualitative attributes, such as Spectral Density of states (DFT)
phase and defect types. Spatial data represents geometries Image Charge density maps (DFT) and stress/
and positional relationships, including crystallographic strain field (FEM)
texture. Table 2 summarizes the common types of data Categorical Phase regions (CALPHAD) and failure
obtained through simulations and experiments. regions (FEM)
Simulation plays a critical role in AI-driven material Spatial Atomic position (DFT)
development by providing high-fidelity datasets that Abbreviations: CALPHAD: Calculation of phase diagrams; DFT:
capture complex material behaviors and properties. These Density functional theory; FEM: Finite element method.
simulations are favorable for generating datasets that
can support predictive modeling and validate AI-driven empirical insights. Common techniques include property
solutions. Commonly employed modeling methods include testing, microscopy, and spectroscopy.
density functional theory (DFT), FEA, and computer
coupling of phase diagrams and thermochemistry In material development for AM, datasets are
(CALPHAD). DFT is widely used for predicting electronic often sourced from literature, online databases, or
43
structures and thermodynamic properties. FEA focuses on experiments. Table 3 summarizes the commonly used
macroscopic phenomena, such as stress-strain behavior online data repositories. While online material databases
and thermal conductivity under various conditions. provide valuable resources for material development, most
CALPHAD, on the other hand, is instrumental in phase existing repositories are primarily built upon data obtained
diagram calculations and thermodynamic modeling. through traditional manufacturing processes. Key aspects
Complementing simulation, experimental measurements of AM, such as the rapid solidification, residual stresses,
provide essential data that capture the real-world behavior and complex microstructural evolution characteristics,
and performance of materials. These measurements not are typically not captured. As a result, applying data from
only validate simulation results but also offer unique these repositories directly to AM may lead to inaccuracies,

Volume 2 Issue 2 (2025) 7 doi: 10.36922/IJAMD025100007