Towards Data-Driven Dynamic Substructuring in Frequency Domain

This repository contains the MATLAB and Python codes for both rectangular plate with matching and non-matching mesh examples from the paper:

M. Faizan Baqir, Hyunwoo Baek, Dahye Son, Peter Persson, and Jin-Gyun Kim
"Towards data-driven dynamic substructuring in frequency domain."

---

Overview

Dynamic substructuring is widely used in structural dynamics to analyze large and complex systems efficiently.
In this work, we integrate a machine learning (ML)-based meta-model with a conventional finite element (FE) model to achieve a hybrid FE–ML framework for vibration analysis in the frequency domain.

Key contributions of this work:

• Replacement of a full FE substructure with a deep neural network (DNN) surrogate model
• Formulation based on dual assembly in the frequency domain
• Handling of non-matching interface meshes via localized Lagrange multipliers

---

How to Run

The workflow is the same for both matching and non-matching mesh examples.

1. Generate datasets (MATLAB)

   • Go to the Data generation/ folder

   • Run: Main_Make_datasets.m

   • This creates training, validation, and test datasets (.csv)

2. Hyperparameter tuning (Python/Jupyter)

   • Run hyperparameters_search.ipynb
   • Select optimal hyperparameters for the generated datasets

3. Model training & prediction (Python/Jupyter)

   • Run DNN_training_Code.ipynb
   • Trains the DNN model and generates predictions

4. Hybrid FE–DNN evaluation (MATLAB)

   • Copy the DNN predictions into the Solution of problem/ folder

   • Run: Main_eval_hybrid_fe_dnn.m

   • Evaluates the FE+DNN hybrid model.

Contact

For questions or discussions, please contact:

Muhammad Faizan Baqir
Ph.D. Student, Modeling and Simulation (M&S) Lab
Department of Mechanical Engineering, Kyung Hee University, South Korea

[email protected]