Physics-informed neural networks for truss FEM calibration using synthetic data: A pratt bridge case study

Details

Supervisors

Saraiva Esteves Pacheco De Almeida, João ; Bal, Ihsan

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

enThis thesis addresses finite element model (FEM) calibration for truss structures under noisy and sparse measurements by framing the task as a supervised inverse problem: recovering member axial rigidities EA from nodal displacements and applied loads. Focusing on an 8-panel, 60m Pratt truss bridge, the work targets non-destructive assessment scenarios with limited high-quality data. A fully synthetic training pipeline was developed using OpenSeesPy to generate large, standardised datasets with controlled parameter ranges and noise models using Sobol sequences to generate space-filling sampling. The case-study geometry, loading, and material bounds reflect typical bridge practice and Eurocode-motivated limits, ensuring physical realism. Two approaches are compared: a purely data-driven, MSE-trained multilayer perceptron and a physics-informed neural network (PINN). On the member-wise prediction task, the PINN consistently outperforms the MSE model but displays unsatisfactory results. Using a category-based formulation, that is, predicting axial rigidity EA for groups of members instead of each member, drastically improves prediction accuracy, showing the PINN retaining superior performance and robustness. Overall, embedding physics as a stabilising prior and adopting design-informed targets yields more reliable, data-efficient calibration. The contributions are (i) an inverse-problem formulation for truss FEM calibration and (ii) a modular synthetic-data generation framework enabling rigorous, reproducible evaluation of machine learning based calibration methods.