Dr.-Ing. | Computational Mechanics | AI in Engineering Science | PINNs, FEM, Data Science & Computer Vision 🚀
Passionate about developing and applying advanced computational methods and AI to solve complex engineering and scientific challenges.
🛠️ My Tech Stack & Expertise
Languages & Scripting:
|
Machine Learning & AI:
|
Computational Science & Simulation:
|
Tools & Platforms:
|
🚀 Highlighted Projects Details for each project can be expanded by clicking on them.
1. Hybrid ML-FEM Viscoelastic-Viscoplastic Damage Model (Published in CMAE)
- Objective: Implemented a sophisticated finite element model combining traditional constitutive laws with LSTM neural networks to simulate complex material behavior in epoxy nanocomposites under cyclic loading, considering moisture and nanoparticle effects.
- Tech Stack: C++, deal.II (FEM Library), Python (for ML aspects), MPI, CMake.
- Key Contributions: Developed a hybrid ML-FEM framework for large deformation solid mechanics, integrated LSTM for computational acceleration, modeled multi-network viscoelastic-viscoplasticity with damage, and incorporated environmental effects.
- Outcome/Impact: Created a robust model for simulating advanced material responses, published in Computer Methods in Applied Mechanics and Engineering. Showcased seamless switching between physics-based and ML models.
- DOI: 10.1016/j.cma.2023.116293
- Repository: [Link to Project Repository]
2. Deep Learning-Based Thermodynamically Consistent Material Model (Published in CMAE)
- Objective: Proposed a physics-informed deep learning (DL) constitutive model for epoxy composites that enforces thermodynamic principles, using experimental data to predict material behavior under diverse ambient conditions (temperature, moisture, nanoparticle volume fraction).
- Tech Stack: Python, Deep Learning (LSTM, Feed-Forward Neural Networks), Experimental Data Analysis.
- Key Contributions: Designed a DL architecture combining LSTM and FFNNs to predict internal variables and free-energy, ensuring thermodynamic consistency. Trained solely on experimental data to capture complex, nonlinear, temperature- and moisture-dependent responses.
- Outcome/Impact: Developed a novel DL model capable of accurately predicting material behavior while adhering to thermodynamic laws, published in Computer Methods in Applied Mechanics and Engineering.
- DOI: 10.1016/j.cma.2024.117038
- Repository: [Link to Project Repository]
3. Physics-informed neural network (PINN) for Hyperelastic Solid Mechanics
- Objective: The project simulates a quasi-static tensile test on a hyperelastic dog-bone specimen, showcasing a powerful, mesh-free approach to solving complex non-linear partial differential equations (PDEs) in computational mechanics.
- Tech Stack: Python, Deep Learning (LSTM, Feed-Forward Neural Networks), Experimental Data Analysis.
- Key Contributions: The core methodology transforms the PDE problem into an optimization task. A neural network is trained to approximate the displacement field, and the loss function is engineered to penalize any violation of the underlying physical laws and boundary conditions.The simulation solves the static equilibrium equation in its strong form for a dogbone specimen. The model is verified using FEM simulations
- Outcome/Impact: Developed a PINN model to solve the strong form for mechanical equilibrium for a dogbone specimen and verified against finite elements simulations.
- Repository: [Link to Project Repository]
4. Variational Physics-Informed Neural Network (VPINN) for 2D Elasticity
- Objective: To solve a 2D linear elasticity problem (plane stress) for a square plate under uniaxial tension. This project demonstrates how VPINNs can solve the weak form of PDEs without traditional FEM meshing.
- Tech Stack: Python, PyTorch, NumPy, SciPy, Matplotlib.
- Key Contributions:
- Implemented a VPINN that transforms the PDE into an energy minimization problem based on the variational (weak) form of the equilibrium equation.
- Utilized Legendre polynomials as test functions and Gauss-Legendre quadrature for numerical integration.
- Enforced Dirichlet boundary conditions analytically using an Augmented Deep Formulation (ADF), ensuring exact satisfaction of boundary constraints.
- Outcome/Impact: Developed a robust VPINN model that accurately predicts displacement, stress, and strain fields. The model's performance was successfully verified against the known analytical solution for the problem.
- Repository: [Link to Project Repository]
5. GNN Force Field for Coarse-Grained Molecular Dynamics
- Objective: To develop a deep learning model capable of predicting atomic forces in coarse-grained molecular dynamics (MD) simulations, creating a fast and accurate surrogate for traditional, computationally expensive force fields.
- Tech Stack: Python, PyTorch, PyTorch Geometric, Pandas, NumPy, Scikit-learn.
- Key Contributions: Implemented a Graph Neural Network (GNN) inspired by the state-of-the-art GNNFF architecture. Engineered a rotationally covariant force prediction module that learns scalar force magnitudes and projects them onto direction vectors, ensuring physical consistency. Developed a robust data pipeline including dynamic outlier filtering and Z-score normalization to handle noisy simulation data.
- Outcome/Impact: Created a complete, end-to-end pipeline for training a GNN-based force field that can significantly accelerate MD simulations of polymer systems, bridging the gap between the speed of classical potentials and the accuracy of ab initio methods.
- Repository: gnn-cg-peo-forcefield
6. Crack Detection in Electromechanical Materials using U-Net (Computer Vision)
- Objective: Applied deep learning (U-Net with ResNet backbones) for semantic segmentation of crack propagation in materials under electromechanical stress, analyzing phase-field and electrical potential patterns from FEM simulations.
- Tech Stack: Python, TensorFlow, Keras, OpenCV, ABAQUS (for data generation).
- Key Contributions: Implemented a multi-class semantic segmentation pipeline for pixel-level crack detection, utilized transfer learning, and automated hyperparameter tuning. Compared phase-field and electrical potential visualization methods.
- Outcome/Impact: Achieved high precision (IoU > 0.95) in detecting and classifying cracks/defects, offering a significant improvement over traditional methods.
- Repository: Computer-Vision-Crack-Detection
7. Predictive Maintenance System for Manufacturing using machine learning
- Objective: Developed a machine learning system to predict 5 different types of equipment failures (TWF, HDF, PWF, OSF, No Failure) in manufacturing environments using sensor data, enabling proactive maintenance.
- Tech Stack: Python, Scikit-learn, XGBoost, Pandas, Matplotlib, Seaborn.
- Key Contributions: Performed comprehensive EDA, extensive feature engineering, implemented multi-class classification models, and created a configurable pipeline with advanced visualizations. Utilized the AI4I 2020 Predictive Maintenance Dataset.
- Outcome/Impact: Built a system achieving strong predictive performance across various failure types, providing a practical solution for reducing downtime in industrial settings.
- Repository: Predictive_Maintenance
8. ABAQUS Multiphysics Diffusion UEL
- Objective: Implemented a User Element (UEL) for ABAQUS to simulate coupled hydro-mechanical behavior of moisture diffusion in polymer materials, capturing stress-assisted transport mechanisms.
- Tech Stack: Fortran, ABAQUS (UEL Development), MATLAB (for visualization).
- Key Contributions: Developed a 20-node quadratic hexahedral UEL for multiphysics coupling (stress-assisted diffusion), enabling monolithic solution of mechanical and diffusion fields. Provided MATLAB tools for post-processing.
- Outcome/Impact: Created a flexible tool for advanced simulation of moisture diffusion in polymers under mechanical stress, applicable to aerospace, marine, and electronics industries.
- Repository: ABAQUS-Multiphysics-Diffusion-UEL
📊 My GitHub Stats
📈 Contribution Graph
📫 How to Reach Me
- Email:
[email protected] - LinkedIn: Dr.-Ing. Betim Bahtiri
- Feel free to open an issue on any of my repositories if you have questions or want to collaborate!
- The projects are not related to my current employer !
Thanks for stopping by! Connect with me to explore collaborations and innovative ideas.