Selected Projects
Some of my recent works. GitHub links in each card.
First/Principle Studies
Fe–Cr Elastic Constants — DFT → ML → Calphad → FEM
Composition dependent Fe–Cr reduced elastic-constant across the full BCC composition range and room temp- to 1200k— corrected Eᵣ tracks Hill VRH within ±12%.
QE · CCX · ParaViewMLP R²=0.91 C11: 312→448 GPaEᵣ: 197→274 GPa
Si Nanowire Mechanical Response — Parametric MD
Uniaxial tensile deformation of [001] Si nanowires across 16 (diameter × strain-rate) combinations using the Tersoff potential. Surface-to-volume ratio, not strain rate, controls stiffness.
LAMMPSTersoff E: 34→44 GPa16-run sweep
Graphene Dirac Cone under Uniaxial Strain — DFT
First-principles PBE study (Quantum ESPRESSO) of Dirac cone robustness under 0–20% zigzag strain. 2D k-scans reveal the 1D band-path overestimates the gap 23× at 4% — the cone survives until ~10%.
Quantum ESPRESSOPBE/PAW 11 strain levels2D k-scan
ML for Materials
EMA-GNN — Structural GNN for Crystal Stability
Structural GNN for direct crystal stability prediction, live on the Matbench Discovery leaderboard — 3rd by F1, 1st by MAE among IS2E direct-prediction models.
PyTorch GeometricMatbench Discovery F1 0.566 · MAE 0.084 eV/atom3rd/6 IS2E · 1st MAE
2D Linear Elasticity via PINN
Physics-informed NN for fast, physics-consistent field predictions.
PINNFEM val loss 2.22
Melting-Point Prediction (2-Level Ensemble)
Stacked RF/XGB/LGBM/MLP with SHAP-guided features.
R² ≈ 0.83~3.041k samples 2-level ensemble
Oxidation-State Assignment
Soft-voting ensemble on lean OQMD/ICSD data — competitive quality on ~15× less data than typical literature sets.
Proxy R² ≈ 0.91~7k OQMD Weighted soft voting
Data & Informatics Infrastructure
Materials Database App
Auto-ETL + Flask UI tracking datasets → features → results with FAIR-aligned provenance and secure file delivery.
Auto-ETLFlask + SQLite FAIR / provenance
Semantic Band-Gap (LLM + KG)
RDF/SPARQL schema with local LLM (llama3.2) for explainable, natural-language materials lookups.
RDF/SPARQLLLM Explainable