2026-09-20
MatSemNet fuses text, numbers, and reaction paths from 700+ papers; grouped-split test R² is 0.727, and La-Co3O4 hits 85.5% ammonia FE in alkali.
Faradaic efficiency (FE) in electrocatalysis is glued to experimental context: composition, electrolyte, pH, potential, and cell type. Pull numbers out of papers and split train/test at random, and near-duplicate conditions leak across the cut. A random split here left training and test samples more than 58% similar on average. The model looks accurate while memorizing lab setups.
Descriptor-only machine learning cannot ingest prose or reported reaction paths. A large language model used alone is too brittle for complex chemistry. Material Semantics Network (MatSemNet), from Peking University, Beijing Technology and Business University, and Sun Yat-sen University Shenzhen, joins the two: an LLM extracts structured records from 700-plus electrocatalysis papers, then a single network encodes text, numeric descriptors, and literature-reported pathways to predict FE for nitrate reduction to ammonia (NO3RR).
Three stages.
For screening they freeze CoOx and sweep rare-earth dopants. SHAP ranks electrolyte as the top feature, matching NO3RR intuition. Several RE elements appear in the training set, but no exact RE-doped spinel Co3O4 recipe matches the later experiments, so the ranking step is already an out-of-formula extrapolation.
On the grouped test split, MatSemNet reaches R² 0.727 and MAE 0.0891 (train R² 0.8203, MAE 0.0225). Classical models on the same split trail on R²:
| Model | Test R² | Test MAE |
| MatSemNet | 0.7270 | 0.0891 |
| Random Forest | 0.3147 | 0.0687 |
| XGBoost | 0.2611 | 0.0727 |
| LightGBM | 0.1820 | 0.0766 |
Tree models can match MAE and still collapse variance, which is why they rank poorly. Encoder swaps: MaterialsBERT R² is 219.1% above SciBERT, 9.16% above MatBERT, and 10.52% above text-embedding-3-large. Adding ReactionLSTM lifts the MaterialsBERT variant by another 31.56% R².
Across acid, alkali, and neutral settings the model ranks lanthanum first. The alkaline order is La > Ce > Pr > Nd > Sm > Gd > Eu. Seven RE-doped Co3O4 nanorods were synthesized. Alkaline measurements track that order, with Sm, Gd, and Eu nearly tied. Best-in-series La-Co3O4 reaches 85.5% ammonia FE and 28.3 mg h⁻¹ cm⁻² at −0.5 V vs. RHE in alkali, with small drift over 30 chronoamperometry cycles. In acid the model quotes high theoretical FE while hydrogen evolution steals current; the paper treats acid numbers as high-uncertainty ranks only.
ICP-OES puts actual RE loadings in a tight 1.14–1.28 wt% band, so dose cannot explain the ranking. DFT finds the strongest NO3⁻ adsorption on La-Co3O4, in the same direction as model and experiment.
The useful object is the loop: literature extraction, leakage-aware evaluation, model ranking, wet-lab follow-up. For anyone training on papers, the 58% similarity check is the takeaway that travels: literature data are not i.i.d., and a random split inflates scores. Treating reported pathways as a context encoder, rather than as discovered mechanisms, is also more honest than the usual "AI found the mechanism" claim.
Reproduction is gated. Code and data are withheld as ongoing work, available only on request. The LLM is unnamed. Anyone hoping to rerun the extraction stack will stall.
The authors position the model as a ranking tool, and they say so for the acidic regime. Validation stays on CoOx with rare-earth dopants, even though the discussion calls the framework material-agnostic. Training saw La and Ce as elements but not this spinel recipe, so the hit is a real extrapolation that still needs a second chemistry.
Extraction accuracy is document-level: one bad key field fails the whole paper. Field-level error rates are missing. Three hundred audited papers against a 700-plus corpus leave residual noise after correction unmeasured. Tree models win MAE and lose R², which means the headline metric was chosen for ranking; calibrated FE values are not the strength.