Graduate Internship at ORNL
May 2024 - Dec 2024; May 2025 - present, Oak Ridge, TN
Summary: At ORNL, I worked on reducing the computational cost of microstructure-sensitive materials modeling by building graph neural network surrogates for composites and polycrystals. My work focused on making structure-to-property prediction more scalable, more robust to heterogeneous microstructures, and more useful for downstream fatigue and damage studies.
Mentor: Massimiliano Lupo Pasini, Computational Coupled Physics, Computational Sciences and Engineering Division (CSED), Computing and Computational Sciences Directorate (CCSD), Oak Ridge National Laboratory
Key outcomes:
- Published paper on graph neural networks for 2D fiber-composite property prediction.
- Built multitask graph surrogates for ferrite-martensite polycrystals to accelerate elastoplastic and fatigue-relevant studies.
- Improved accuracy and efficiency relative to CNN baselines, including stiffness prediction with 160x fewer parameters and peak-strength prediction with better accuracy using 12x fewer parameters.
- Contributed to HydraGNN code quality and scientific machine learning workflows on Linux/HPC systems.
Project Highlights
1. Fiber-composite surrogate modeling
Situation: High-fidelity homogenization of 2D fiber composites is expensive, especially when microstructures are statistically diverse, anisotropic at the SVE scale, and far from the RVE limit.
Task: Build a surrogate that can predict stiffness and strength directly from microstructure while remaining accurate for high material-contrast cases and practical for large parametric studies.
Action: I developed topology-based GNN workflows for 2D fiber composites, trained them to predict the full stiffness tensor together with peak-strength and brittle-fracture-related quantities, introduced physics-based normalization for extreme contrast ratios, and used Voronoi-derived features to improve learning in small-data regimes.
Result: The models remained robust across diverse microstructure configurations, achieved stiffness prediction with 160x fewer parameters than CNN baselines, improved peak-strength accuracy while using 12x fewer parameters, and led to a published paper in Materials & Design.
Graphical abstract (Caliskan et al., 2025, M&D).
2. Multitask polycrystal surrogate modeling
Situation: Elastoplastic response prediction for polycrystals typically requires large CPFEM ensembles, which becomes a bottleneck when studying variability, extreme responses, and fatigue-relevant random fields over many statistically distinct SVEs.
Task: Create a surrogate that preserves microstructure-sensitive variability while predicting both scalar elastoplastic measures and full stress-strain behavior across compositions and SVE sizes.
Action: I represented dual-phase ferrite-martensite polycrystals as grain-adjacency graphs with phase, geometry, orientation, and misorientation information, then developed multitask GNN surrogates that jointly predict elastic-plastic quantities of interest and stress-strain responses across martensite volume fractions and SVE sizes.
Result: The workflow supported population-level comparisons of finite-SVE variability and enabled statistically consistent random-field construction for mesoscale fatigue and damage analyses, making large ensemble studies more tractable.
Selected Outputs
- Caliskan, Erdem, Reza Abedi, and Massimiliano Lupo Pasini. “Graph Neural Networks for Mechanical Property Prediction of 2D Fiber Composites.” Materials & Design (2025): 114500. Publication link
- Caliskan, Erdem, Anik Das Anto, Massimiliano Lupo Pasini, Stephanie TerMaath, and Reza Abedi. “Multitask Graph Neural Networks for Elastoplastic Response Prediction in Dual-Phase Polycrystals.” Journal of Materials Science: Materials Theory (under review). Publication link
- Caliskan, Erdem, Anik Das Anto, Reza Abedi, and Massimiliano Lupo Pasini. “Probabilistic Multi-Task Graph Neural Network Surrogates for Elastic-Plastic Behavior and Fatigue Indicator Prediction in Polycrystalline Alloys.” In SES Conference 2025, Atlanta, Georgia, USA, October 12-15, 2025.
Methods/Stack: HydraGNN, PyTorch, PyTorch Geometric, microstructure-derived graphs, grain-adjacency graphs, physics-based normalization, Voronoi-derived features, and Linux/HPC workflows.