SIMPLEX: Composition-Space Deep Learning for Hydrogel Adhesion with Out-of-Distribution Extrapolation to Model-Discovered High-Performance Formulations
SIMPLEX (Simplex composition encoding with Interaction-aware attention, Multi-modal fusion, Pretraining-ready regularisation, Learnable domain constraints and EXtrapolation evaluation) is a dual-modality residual deep network that maps the monomer composition of hydrogels (six molar fractions on the composition simplex plus their 15 pairwise interactions) to underwater glass adhesion strength (kPa).
Key idea — model-guided extrapolation evaluation. A model trained on 180 low-performance formulations must rank 161 high-performance formulations discovered by the original study's sequential model-based optimisation (SMBO) loop — a target-value extrapolation beyond the training range. Under this protocol SIMPLEX significantly outperforms tree ensembles on ranking (external Spearman rho 0.50 vs 0.21 for random forest; paired bootstrap difference +0.18, 95% CI [0.07, 0.31]) and achieves the best top-k screening precision, while matching the strongest baseline in-distribution.
| Metric | SIMPLEX | RandomForest | ElasticNet |
|---|---|---|---|
| Internal CV R² | 0.709 ± 0.077 | 0.719 (tie) | — |
| External Spearman ρ | 0.501 [0.369, 0.619] | 0.211 | 0.494 |
| External Top-20 precision | 0.25 | 0.10 | 0.05 |
sheng-hu/hydrogels(MIT licence), accompanying Liao et al., Nature 644, 89–95 (2025), DOI 10.1038/s41586-025-09269-4.- Internal:
data/raw/df_180.csv(180 round-1 formulations). - External:
data/raw/df_341.csvminusdf_180.csv(161 SMBO-discovered formulations).
cd code
C:/Users/TS/.conda/envs/HydroGelNet/python.exe run_all.py --allStages: download → build → qc → tune → train → baselines → stats → gate → interpret → figures → tables → paper. PERF-GATE must pass before writing.
paper/frontiers_SIMPLEX.tex— Frontiers LaTeX (Overleaf project 6a6a083446657df2cc7a741e).paper/SupplementaryMaterial.md/paper/simplex_supplementary.tex.paper/manuscript.md— Markdown source.FRONTIERS_SUBMISSION.txt— title + article type (Original Research).
Python 3.11 (conda env HydroGelNet, cloned from py311), PyTorch 2.14.0
- cu130, NVIDIA RTX 5080 16 GB. Fixed seeds [42, 2024, 7, 1337, 20260731]; all preprocessing fitted inside training folds only.