ASAMS: An Adaptive Sequential Sampling and Automatic Model Selection for Artificial Intelligence Surrogate Modeling

Area of Science:

• Computational Science and Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Surrogate Modeling (SM) reduces computational costs in system simulations.
• Digital Twins (DT), Design Mining (DM), and Soft Sensors (SS) present new challenges for SM due to complex AI algorithms and limited physical experiments.
• Existing adaptive sampling methods are often limited to Kriging models.

Purpose of the Study:

• To develop an integrated framework for generating artificial intelligence-based surrogate models.
• To minimize system evaluation and maximize model performance.
• To assist in model selection within an automated machine learning (AutoML) context.

Main Methods:

• Integration of a novel adaptive sampling methodology with an AutoML framework.
• Utilizing a grid search algorithm for candidate model selection.
• Employing leave-one-out cross-validation and Voronoi diagrams for efficient sampling and model accuracy estimation.

Main Results:

• The proposed framework effectively reduces the number of candidate models in each iteration.
• The method demonstrates improved accuracy and performance for AI-based surrogate models.
• Validation through two case studies confirms the framework's applicability.

Conclusions:

• The integrated adaptive sampling and AutoML approach offers a powerful solution for developing efficient and accurate AI-based surrogate models.
• This methodology addresses the challenges posed by modern AI-driven simulation techniques.
• The framework provides a systematic way to optimize surrogate model performance with minimal experimental data.