Machine learning-based regime identification and size-resolved optimization of charged-particle capture and rebound on spherical collectors

Abstract

Particle rebound at the collector surface represents a critical loss mechanism in electrostatic precipitation systems, where operating conditions that maximize collision efficiency often increase the rebound fraction, thereby limiting overall collection efficiency. Balancing these competing objectives across a high-dimensional design space remains a significant engineering challenge.

This work presents a clustering-guided surrogate-assisted optimization framework that integrates surrogate modeling, K-means clustering, Non-dominated Sorting Genetic Algorithm II (NSGA-II), multi-objective optimization, and local explainable artificial intelligence (XAI). From an artificial intelligence perspective, the principal contribution is a data-driven strategy for optimization bound selection through clustering-guided design-space reduction, replacing manually defined optimization bounds with an automated and physically meaningful alternative.

From an engineering perspective, the proposed framework optimizes charged particle collection on spherical collectors while minimizing rebound-induced losses across a polydisperse particle size range. A LightGBM surrogate model trained using approximately 2,100 high-fidelity finite-element simulation samples predicts system behavior over nearly 5 × 10⁵ parameter combinations. K-means clustering identifies favorable operating regions whose statistical bounds constrain the NSGA-II search space, significantly improving optimization efficiency.

Pareto-optimal trade-offs between collision efficiency and rebound fraction are obtained for particle diameters ranging from 0.5 to 10 μm, with the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) employed for selecting the final optimal solutions. The results reveal a particle-size-dependent transition from electrostatically dominated transport for small particles, where high collection voltages achieve collection efficiencies exceeding 98% with rebound below 5%, to inertia-dominated transport for larger particles, where lower voltages maintain collection efficiencies above 84%.

Local explainable artificial intelligence analysis identifies flow velocity as the dominant parameter governing both particle collection and rebound suppression across all particle-size regimes. The proposed framework demonstrates how machine learning, optimization, and explainable artificial intelligence can be integrated with high-fidelity physics-based simulations to develop efficient and physically interpretable engineering design methodologies.

Keywords

• Scientific Machine Learning
• Surrogate Modeling
• Multi-objective Optimization
• NSGA-II
• K-means Clustering
• Explainable Artificial Intelligence
• SHAP
• Electrostatic Precipitators
• Particle Rebound
• TOPSIS

Recommended citation: Srivastava, A., Kandpal, S., Ray, B., Kumar, M., & Dasgupta, D. Machine learning-based regime identification and size-resolved optimization of charged-particle capture and rebound on spherical collectors. Submitted to Powder Technology.
Download Paper | Download Bibtex