Research Outputs

1.Conference Poster in AIMSE, 2023 (22-23 November, Saarbrücken, Germany):: A two-yearly Conference organized by DGM, focused on the application of AI across sub-domains of Materials Science & Engineering. My entry was in Predicting properties/microstructures which demonstrated SINDy's equation discovery correctly depicts a dynaamic system (2 independent magnetostrictive excitation channels) only from time-series data. The dynamics depicted was validated against analytical stiffness behavior (force-deflection curve), damping behavior (phase portraits), potential energy surface (self-excitation hardened vs. linear) of the respective duffing oscillator in consideration. See below for a deeper peek.

2.Neuro-Symbolic Knowledge Graph for Semiconductor Band-Gap Screening, deployed 2026:: A fine-tuned Llama 3.2 (3B) LoRA adapter translates natural-language questions into SPARQL, executed against an RDF graph of 150,987 Materials Project entries (1.36M triples). A deterministic sanitizer sits between model and graph, repairing and schema-enforcing every generated query rather than trusting raw model output. The graph enables single-query, symmetry-aware screening — e.g. non-centrosymmetric materials with a band gap above 2.5 eV, a necessary-condition screen for nonlinear-optical candidates, returning 9,429 of 150,987 materials with provenance attached. Node identity was rebuilt onto the Materials Project material_id after diagnosing that formula-based identity had silently merged 48,569 entries across distinct structural phases; phase distinguishability was independently verified with pymatgen's StructureMatcher (91.6% of multi-entry compositions confirmed structurally distinct). Generalization of the fine-tuned query model was tested explicitly on an element pair absent from all 97 trained combinations, rather than relied on from the loss curve alone. See below for the live interface and full pipeline.

3.EMA-GNN — Structural Graph Neural Network for Formation-Energy Prediction, Matbench Discovery, merged 2026-08-12:: A message-passing GNN predicting formation-energy-per-atom directly from an unrelaxed crystal structure, with no relaxation step at inference. The architecture follows the structural model described in Merchant et al. (GNoME, Nature 2023) as a starting hypothesis rather than an assumption — an independent hyperparameter search over width, depth, activation and learning rate, run on this project's own data, converged on the same configuration the paper reports. Benchmarked on the WBM test set (256,963 structures) against five other direct-prediction (IS2E) models on the leaderboard, EMA-GNN ranks 1st by MAE (0.084 eV/atom) and R² (0.387), and 2nd by F1 (0.566, within 0.002 of the top score) — metrics recomputed and made authoritative by the Matbench Discovery scoring pipeline itself, not self-reported. Two methodological errors were found and corrected during development rather than after: an initial mean-based test-time-augmentation aggregation (the paper specifies minimum), and a naive stability threshold that conflates exothermicity with thermodynamic stability (correcting to hull-displacement scoring moved F1 from ≈0.34 to 0.56 with no change to the model itself). Checkpoints and predictions are archived on Figshare under a permanent DOI.