Adaptive AI-based surrogate modelling via transfer learning for DEM simulation of multi-component segregation.

Area of Science:

• Computational physics and engineering
• Materials science and granular mechanics
• Machine learning applications in scientific modeling

Background:

• Granular material segregation is a significant industrial challenge.
• Discrete Element Method (DEM) simulations provide insights but require extensive calibration.
• Machine learning (ML) surrogate models (SMs) offer a promising solution for DEM calibration.

Purpose of the Study:

• To develop effective SMs linking DEM interaction parameters to granular segregation.
• To evaluate various ML models and optimize hyperparameters using Bayesian optimization.
• To introduce a transfer learning (TL) approach for adaptive SMs in new scenarios.

Main Methods:

• Trained multiple ML models (ANNs, ensemble learning) on cost-effective DEM simulation data.
• Employed Bayesian optimization with cross-validation for hyperparameter tuning.
• Developed a novel TL-based approach using Gaussian process regression (GPR) for unseen scenarios.

Main Results:

• Gaussian process regression (GPR) demonstrated high accuracy with very small datasets.
• The TL approach enabled accurate SMs for unseen initial configurations with few samples.
• Performance improvement of 17% (1 sample) and 47% (5 samples) was observed for TL-GPR on unseen scenarios.

Conclusions:

• The proposed TL-based SMs significantly reduce the data burden for DEM calibration.
• This methodology accelerates the development of reliable DEM models for granular materials.
• The findings facilitate efficient calibration and prediction of granular segregation phenomena.