CdSe-QDs-GNN-Framework

Machine learning framework for predicting time-resolved electronic properties in ligand-passivated CdSe quantum dots (Cd₂₈Se₁₇X₂₂, X = Cl, OH). This repository includes AIMD trajectories, DFT-calculated electronic properties, and graph-based neural network model (ALIGNN), along with atom-specific importance analyses via Feature Nullification Analysis (FNA). Developed for studying bandgap and subgap fluctuations over extended trajectories using transfer learning and minimal DFT sampling.

---

Structure–property evolution

The short clip below shows how the Cd₂₈Se₁₇Cl₂₂ core–ligand geometry (left) evolves together with the ensemble‑predicted bandgap trajectory (right) over the first 10 ps of the AIMD simulation, highlighting the structure–property correlations captured by ALIGNN.

---

Ensemble‑model accuracy 

The animation below compares parity plots for every individual ALIGNN model (run 0 → run 19, left panel) with the ensemble‑averaged parity plot (right panel). A tight 1:1 line on the right demonstrates how averaging 20 models reduces scatter and delivers band‑gap predictions that closely match DFT.

---

Atom‑level importance 

Feature Nullification Analysis (FNA) bins atoms by how much removing their information increases the ensemble RMSE:

Group (meV bin)	Atoms
G 1 30 – 25 meV	Se2, Se1
G 2 25 – 20 meV	Se5, Cl7
G 3 20 – 15 meV	Se16, Cl3, Cl9, Cl10
G 4 15 – 10 meV	Se3, Cl13, Se7, Se11, Se8, Se6, and so on
…	lower‑impact atoms

The GIF couples the bar‑chart of importance scores (left) with the highlighted atom positions inside the Cd₂₈Se₁₇Cl₂₂
structure (right), making it easy to see which sites dominate bandgap fluctuations.

---

Full AIMD trajectories (30 000 files)

To keep this repository small, the complete 15 ps trajectories are not stored in Git. Download the tar archives from the Releases tab:

System	Release asset	Size
Cd₂₈Se₁₇Cl₂₂	Cd28Se17Cl22_15000_vasp.tar.gz	20 MB
Cd₂₈Se₁₇(OH)₂₂	Cd28Se17OH22_15000_vasp.tar.gz	27 MB

---

Zenodo archive (everything in one place)

All numerical artefacts supporting this repository have been deposited on Zenodo:

https://doi.org/10.5281/zenodo.15359153

What you’ll find inside the archive (≈ 4.37 GB):

Category	Contents
AIMD data	15 ps, 1 fs‑step trajectories for Cd₂₈Se₁₇Cl₂₂(*.vasp)
DFT labels	Bandgap values used for ALIGNN training (id_prop.csv)
Ensemble models	20 ALIGNN checkpoints (run_*/temp/checkpoint.pt, 0 – 10 ps training)
Predictions	Per‑frame bandgap for 0 – 10 ps (prediction.csv)
Atom‑importance	Feature Nullification outputs for every model × atom (atom_imp_*)
Transfer‑learning	Fine‑tuned checkpoints + predictions for the extended 10 – 15 ps window
SLURM logs & scripts	All job scripts

Download the archive to reproduce every figure in the manuscript or to kick‑start your own experiments with pre‑trained models.

---