A Metamodel-Based General-Purpose Autocalibration Tool for Simulation Models.

Area of Science:

• Computational Science
• Optimization Methods
• Simulation Modeling

Background:

• Simulation calibration configures model parameters to match observed data, often computationally expensive.
• Metamodels offer a balance between accuracy and efficiency for complex parameter spaces in simulation calibration.

Purpose of the Study:

• To introduce and evaluate novel metamodel-based optimization and hybrid strategies for simulation calibration.
• To compare the calibration accuracy and computational cost of different simulation calibration approaches.

Main Methods:

• Evaluated four simulation calibration methods: Randomly-Simulate (RS), Optimally-Predict (OP), Predict-then-Simulate (PtS), and Simulate-then-Predict (StP).
• Compared methods based on calibration accuracy and computational expense, utilizing metamodel-based optimization and simulation runs.

Main Results:

• The Optimally-Predict (OP) approach reduced computational cost but hybrid strategies showed superior performance.
• The Predict-then-Simulate (PtS) approach achieved a 46% average reduction in total error compared to RS, with maintained computational efficiency.

Conclusions:

• A novel metamodel-based optimization approach enhances simulation calibration for computationally expensive studies.
• The proposed hybrid framework, particularly PtS, offers a balance between accuracy and efficiency, with potential for extension to stochastic models.