An innovative artificial intelligence approach for predicting particle size distribution in Al-B₄C powders using neural networks
Authors/Creators
Description
Highlights of the Study
This study develops and validates an artificial neural network (ANN)-based framework for predicting particle size distribution (PSD) and understanding morphological evolution during mechanical milling of Al–B₄C nanocomposite powders. By integrating ANN modeling with laser particle size analysis, SEM, and surface topography characterization, the study demonstrates the capability of artificial intelligence to reproduce both normal and cumulative PSD curves and predict key particle characteristics under different milling conditions.
Key findings:
🟢 High ANN predictive accuracy: The ANN model demonstrated excellent predictive performance across different Al–B₄C powder compositions and milling conditions, with R² values consistently above 0.98 and RMSE values as low as 4.2.
🟢 Accurate D50 prediction: The ANN successfully tracked the evolution of D50 during mechanical milling. D50 initially increased by approximately 25% during the first 4 h because cold welding dominated the process, followed by an approximately 30% decrease as fragmentation became the prevailing mechanism.
🟢 Cold welding–fragmentation transition: The initial increase in particle size resulted from deformation and cold welding of aluminum particles, whereas prolonged milling induced strain hardening and fragmentation, progressively reducing particle size and ultimately producing a more stable particle population.
🟢 Morphological transformation: SEM observations showed that aluminum particles transformed into flake-like morphologies after approximately 2 h of milling, developed increasingly complex cold-welded structures at intermediate times, and evolved toward refined equiaxed particles after 16 h.
🟢 Bimodal PSD prediction: The ANN successfully identified and reproduced bimodal particle size distributions during the early and intermediate stages of milling, demonstrating its sensitivity to the simultaneous presence of B₄C-rich fine-particle populations and larger aluminum-based particles.
🟢 B₄C incorporation mechanism: The initially distinct B₄C-related PSD peak progressively disappeared with milling. After 16 h, SEM observations indicated that B₄C nanoparticles had become embedded within the aluminum matrix, explaining the disappearance of the separate B₄C peak from the PSD curves.
🟢 FWHM evolution: Full Width at Half Maximum (FWHM) analysis revealed an initial broadening of the particle size distribution during the first 4 h, followed by progressive narrowing as continued fragmentation and cold welding produced a more homogeneous particle population.
🟢 Surface topography evolution: Surface characterization demonstrated substantial changes in roughness, waviness, and isotropy during milling. Overall, surface roughness decreased by approximately 40% as prolonged mechanical impacts and friction progressively smoothed the powder surfaces.
🟢 Improved particle isotropy: Milling initially generated anisotropic plate-like particles, but prolonged impact, fracture, and plastic deformation progressively randomized particle morphology and promoted a more isotropic and equiaxed structure.
🟢 Effect of B₄C content: Increasing B₄C content generally shifted particle size distributions toward finer sizes. The embedded ceramic nanoparticles accelerated work hardening of the aluminum matrix, thereby promoting fragmentation and refinement during milling.
🟢 Effect of B₄C particle size: The ANN successfully captured the influence of B₄C nanoparticle size on PSD for both 21 and 71 μm aluminum powders, demonstrating that the relationship depends on the initial aluminum particle size and the efficiency of B₄C incorporation into the matrix.
🟢 Experimental–ANN agreement: Predicted normal and cumulative PSD curves closely followed the corresponding experimental distributions, confirming that the ANN could reproduce complex nonlinear relationships among milling time, aluminum particle size, B₄C size, B₄C content, and resulting powder characteristics.
🟢 Process optimization potential: The developed ANN framework provides a computationally efficient approach for predicting powder behavior and optimizing mechanical milling conditions, with potential future applications in real-time process monitoring, digital twins, intelligent manufacturing, powder metallurgy, and additive manufacturing.
Overall, this study demonstrates that artificial neural networks can accurately predict not only conventional particle-size metrics such as D50, but also complete normal and cumulative PSD curves, FWHM evolution, and complex bimodal distributions. The strong agreement between ANN predictions and experimental observations establishes AI-assisted modeling as a powerful tool for understanding and optimizing the mechanical milling of Al–B₄C nanocomposites.
KEYWORDS
Artificial Neural Network (ANN) • Particle Size Distribution (PSD) • Al–B₄C Nanocomposites • Mechanical Milling • D50 • Full Width at Half Maximum (FWHM) • Bimodal Distribution • Machine Learning • Powder Morphology • Nanoparticles
How to cite
Khakbiz, M., Akhlaghi, F., & Rezaee-Bazzaz, A. (2025). An innovative artificial intelligence approach for predicting particle size distribution in Al-B₄C powders using neural networks. Materials Today Communications, 49, 113746.
DOI: 10.1016/j.mtcomm.2025.113746
Abstract
Abstract
This study aims to develop and validate a neural network-based modeling approach to predict particle size distribution characteristics and understand the evolution of powder morphology during mechanical milling of Al-B₄C nanocomposites. The ANN model was developed to simulate and predict both normal and cumulative PSD curves with high accuracy, achieving a correlation coefficient (R² > 0.98) across various powder mixtures.
By adjusting milling parameters, such as B₄C content (5 % and 10 %) and nanoparticle sizes (90 nm, 700 nm, 1200 nm), the model accurately predicted key characteristics like D50, which initially increased by 25 % due to cold welding in the first 4 h, followed by a 30 % reduction as fragmentation became the dominant mechanism.
The ANN also effectively captured the bimodal particle size distributions, providing valuable insights into the interaction mechanisms between aluminum and B₄C particles during the milling process. In parallel, SEM analysis provided visual confirmation of the milling process, revealing the transformation of aluminum particles from flake-like morphologies after 2 h to refined, equiaxed structures after 16 h of milling.
The topography analysis showed a progressive 40 % reduction in surface roughness and a marked decrease in waviness, indicating the smoothening of powder surfaces over time. These findings were supported by Full Width at Half Maximum (FWHM) measurements, which indicated a broadening of the PSD in the early stages of milling, followed by a shift toward a more uniform distribution with increased milling time.
Together, the ANN modeling and experimental results provide a comprehensive understanding of the mechanical milling process, offering a powerful approach for optimizing the fabrication of high-performance Al-B₄C nanocomposites.