AI-enhanced support vector machine framework for nanoparticle size and surface nanotopography analysis--颗粒学报PARTICUOLOGY

Khakbiz, M., Shahmoradi, M. R., Akhlaghi, F., & Soroush, K. (2025). AI-enhanced support vector machine framework for nanoparticle size and surface nanotopography analysis. Particuology, 106, 156-173. https://doi.org/10.1016/j.partic.2025.07.017

Mehrdad Khakbiz^a *, Mohammad R. Shahmoradi^b, Farshad Akhlaghi^c, Kimia Soroush^d

a Department of Chemical and Biochemical Engineering, Rutgers, The State University of New Jersey, Piscataway, NJ, USA
b College of Interdisciplinary Science, University of Tehran, Tehran, Iran
c School of Metallurgy and Materials Engineering, College of Engineering, University of Tehran, Iran
d School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran

10.1016/j.partic.2025.07.017

Volume 106, November 2025, Pages 156-173

Received 6 May 2025, Revised 4 July 2025, Accepted 7 July 2025, Available online 28 July 2025, Version of Record 13 September 2025.

E-mail: Mehrdad.Khakbiz@rutgers.edu

Highlights

• SVM models achieved R² up to 0.99 for accurate particle size distribution prediction.

• SEM analysis revealed morphological transitions during Al-B₄C powder milling.

• Surface roughness increased to 0.39 nm,analyzed using SPIP™ and GPR models.

• Mid-sized B₄C particles showed optimal reduction with balanced milling times.

• Increasing B₄C content to 10 % improved SVM accuracy in capturing milling dynamics.

Abstract

This study presents a machine learning-based approach using Support Vector Machines (SVM) to model the particle size distribution (PSD) and predict surface characteristics of Al-B₄C nanocomposite powders synthesized through high-energy ball milling. Two SVM kernels, Polynomial and Radial Basis Function (RBF), were applied to simulate PSD curves and surface morphology, with experimental validation conducted via laser particle size analysis and scanning electron microscopy (SEM). The models demonstrated strong predictive capabilities, achieving R² values between 0.91 and 0.99 and cross-validation coefficients (q²) from 0.93 to 0.99. Normal distribution models yielded lower RMSE values (0.11–2.13) compared to cumulative distribution models (4.34–6.55), indicating higher precision in modeling. SEM analysis revealed morphological transitions during milling, with particles evolving from spherical to fragmented shapes after 4 h. Surface metrics including roughness, waviness, and isotropy were quantified, showing that isotropy decreased from 82.48 % at 0 h to 57.69 % at 4 h due to directional deformation, then partially recovered to 62.50 % at 10 h. Gaussian Process Regression (GPR) showed strong alignment with experimental surface trends and accurately predicted nanoscale topographic parameters. Response Surface Methodology (RSM) was employed to visualize size reduction behavior for B₄C particles with initial sizes of 90, 700, and 1200 nm. For 700 nm particles, size reduction stabilized beyond 10 h, while 90 nm particles exhibited rapid refinement within the first 5–10 h. In contrast, 1200 nm particles showed slower, continuous reduction requiring >15 h of milling. SVM models successfully captured these nonlinear trends,with minor underestimations at intermediate time points. RSM plots for aluminum particle sizes (21 and 71 μm) revealed that Al-21 led to stable and uniform distributions,whereas Al-71 exhibited nonlinear behavior with volume percentage drops under specific conditions. These findings confirm that SVM and GPR are robust tools for modeling PSD and surface evolution in ball-milled nanoparticles.

Graphical abstract

Keywords

Support vector machine (SVM); Particle size distribution curves; Kernels; Nanocomposites; Nanocrystalline; Gaussian process regression (GPR)

文章导读