A probabilistic metric for the validation of computational models

Area of Science:

• Computational modeling and simulation
• Data analysis and validation
• Structural analysis

Background:

• Traditional validation metrics often struggle with large data variations and high-dimensional datasets.
• Assessing the real-world representativeness of model predictions requires robust quantitative methods.
• Existing relative error metrics are typically limited to sequential data, hindering application to complex data fields.

Purpose of the Study:

• To introduce a new, robust validation metric for assessing model prediction accuracy.
• To enable the application of validation metrics to high-dimensional data fields using dimensionality reduction.
• To provide a quantitative measure of model-data agreement under specific experimental conditions.

Main Methods:

• Development of a novel validation metric combining uncertainty thresholds and normalized relative error.
• Application of orthogonal decomposition to reduce data matrix dimensionality into feature vectors.
• Validation of the metric using three established case studies in structural analysis.

Main Results:

• The proposed metric provides a probability of model predictions representing real-world data.
• The metric demonstrates robustness against significant variations within the measurement data.
• Dimensionality reduction successfully adapted the metric for use with data fields.

Conclusions:

• The new validation metric offers a reliable method for assessing model performance across various scientific fields.
• The technique is adaptable to diverse disciplines and sectors relying on modeling and simulation.
• This quantitative approach enhances the rigor of the model validation process.