Md. Saidul Islam

Themes

• Source Nonlinearity of implantable Electromagnetic sensors:Sensor-Coil sytem + MATLAB + Audio Interface -based excitation/acquisition pipeline with SINDy-discovered Duffing dynamics (interpretable signal + derivatives) validated via energy, stiffness, and damping analyses; ANN models (LSTM/MLP) used for comparative prediction R² ≈ 0.99 (SINDy) on acquired signals.
• Oxidation-state assignment (Reproduction & Extension): A soft-voting tree-based ensemble (GB, ET, RF, LGBM) with hybrid hyperparameter tuning on ~7k OQMD samples, reached R² ≈ 0.91 (proxy metric), delivering competitive quality on ~15× less data than typical literature sets.
• Melting-point prediction (Reproduction & Extension): A 2-level ensemble with decorrelated-custom-stacked (RF/LGBM/MLP) base learners and meta learner (GB) on ~3.041k record, achieved R² ≈ 0.83 s.
• High-throughput semantic Knowledge-Graph (Bandgaps of Semiconductors): RDF/SPARQL schema parsed (primary task)/queried (secondary task) via a local LLM (llama3.2:3b) for explainable lookups.
• High-throughput-style materials database app Flask/Jinja + SQLite/Drive backend with ETL pipelines and admin CRUD, enabling FAIR-aligned documentation (raw → features (own projects) → results) with provenance tracking and secure file delivery.