Experimental Validation of Surrogate Machine Learning Models for Finite Elements Applications

Exposure to unexpected loads, aging material, and environmental conditions can trig-ger significant risks to the structural integrity of key infrastructures like bridges, buildings, and aircraft. Structural Health Monitoring approaches can be employed to understand and assess the condition of the structures in real time, helping detect damage and guide maintanance decisions. One of the biggest limitations of Structural Health Monitoring approaches is in predicting how a structure would respond under untested extreme loads. Conducting either physical testing or Finite Element Analysis for every possible scenario is time consuming and computationally expensive. The so-lution to this problem lies in surrogate finite element models, in which finite elements models are combined with machine learning approach to quickly and accurately evaluate structural vibrations for a wide range of conditions. Machine learning algo-rithms, capable of processing large datasets and identifying complex patterns, can be trained to model the behavior of structures subjected to time-varying loads. This study presents the experimental validation of surrogate finite element models constructed using artificial neural networks. These surrogate models provide mid-fidelity estimates of structural responses at arbitrary time instances and spatial locations during opera-tion. By learning the dynamic behavior of the structure, the surrogate is employed to develop a digital twin of the system. Experimental results confirm that the surrogate model improves the predictive accuracy of the baseline finite element model by incor-porating measurement data, thereby offering a robust representation of structural be-havior.