COSMOBridge — data, frozen predictions and DFT outputs

Archive accompanying "COSMOBridge: Bridging 2D Molecular Graphs and 3D Electrostatic Surfaces for Ionic Liquid Thermodynamic Property Prediction" (submitted to the Journal of Chemical Information and Modeling).

Source code: https://github.com/taminulislam/COSMOBridge (MIT licence). Archive DOI: 10.57967/hf/11424. This repository is the permanent record of the data, models and predictions behind the paper's tables and figures.

The predictions in predictions/ are deposited before the validation measurements described in Section 5.3 of the paper are made. Their purpose is to fix the model's output so that those measurements score predictions made ex ante. Please do not amend that directory after the measurements begin.

Contents

Path	What it is	Used by
dataset/	the curated 28-ionic-liquid dataset, split by ionic-liquid identity into 19 train (152 rows) / 4 validation (32) / 5 test (39), plus split_info.json naming the ILs in each split	Table 1, §2.2, all training
predictions/candidates_110_frozen_ranking.csv	the full 110-candidate virtual screen, per-seed means and standard deviations, in original units	Table 7, Figure 7
predictions/test_set_predictions.npz	held-out test-set predictions: 5 COSMOBridge seeds, the Chemprop baseline, the fusion path, and the targets	Tables 3 and 4, the bootstrap tests
predictions/top20_candidates.csv	the 20 candidates carried forward to DFT	SI Table S2
dft/nwchem_outputs/	raw NWChem B3LYP/def2-SVP + COSMO outputs, 40 single points covering 16 unique ions	SI Table S2
dft/nwchem_inputs/, dft/geometries/	the inputs and GFN2-xTB geometries those runs used	reproduction
checkpoints/	trained weights for the reported models	Table 3
analysis/*.json	outputs of the bootstrap and seed-matched baseline runs, as reported in the paper	§4.1, Table 3
experimental/	characterization report and figures for the two synthesized salts	§5.2, SI §S2

Reproducing the analyses

All code lives in the GitHub repository above; this archive holds only data and model weights. The scripts scripts/revision/b5_chemprop_multiseed.py (Chemprop over the five shared seeds, Table 3) and scripts/revision/b6_cluster_bootstrap.py (row-level and IL-level bootstrap, Section 4.1) read dataset/ and predictions/ from here and reproduce analysis/b5_chemprop_multiseed.json and analysis/b6_cluster_bootstrap.json. analysis/c1_official_bootstrap.json is the original row-level bootstrap that the paper reports alongside them.

Notes on the data

• The 39 test rows are 8 temperatures of each of 5 ionic liquids and are not independent. b6_cluster_bootstrap.py resamples at both levels; the IL-level interval is the conservative one.
• The DFT solvation energies are a null result. They are the (electrostatic) solvation energy line of each NWChem COSMO block, reported as E_sol - E_gas, and they do not correlate with the screen's ranking (Spearman rho = -0.28, p = 0.23). Do not read them as support for the ranking.
• Activity coefficients are plain gamma, not ln gamma, at x1 = 0.5 and the stated temperature. The identity G^E = RT(x1 ln g1 + x2 ln g2) closes to within 0.09 kcal/mol on this footing.
• The dataset grid is complete except for EMIMOAc at 323.15 K: 223 rows, not 224.
• No data are withheld.

Citation

@article{cosmobridge2026,
  author  = {Ahmed, Khaled and Islam, Taminul and Acquah, Alex and Achinivu-Ibagere, Ezinne C.},
  title   = {{COSMOBridge}: Bridging {2D} Molecular Graphs and {3D} Electrostatic
             Surfaces for Ionic Liquid Thermodynamic Property Prediction},
  journal = {Journal of Chemical Information and Modeling},
  year    = {2026},
  note    = {Submitted}
}