MatSemNet: a literature-trained model ranks seven rare-earth dopants, and the lab measures the same top four

On 15 September 2026 a team from Peking University, Beijing Technology and Business University and Sun Yat-sen University published MatSemNet (Material Semantics Network) in npj Computational Materials. The version online is the peer-reviewed, accepted article preview, which the publisher says will be replaced by the final Version of Record. The question is whether electrocatalysis data scattered through papers, under very different experimental conditions, can train a model good enough to screen catalysts.^[S1]

The data: extracted by an LLM, checked by people

A few-shot prompt had an LLM extract catalyst type, metals, architecture, defect engineering, cell type, electrolyte, nitrate concentration, pH, optimum voltage, stability, reaction pathway, adsorption and activation energies and Faradaic efficiency from more than 700 electrocatalysis papers, as XML. In a check of 300 papers in three batches of 100, 12, 14 and 20 papers carried extraction errors, which the paper reports as roughly 80-90% document-level accuracy before correction. Every record used for training was verified by hand against the source, and papers with unresolved ambiguities were dropped. The paper does not state how many records the final dataset holds.^[S1][S2]

The statistics show Co-based catalysts, CoOx above all, dominating the literature, alkaline media as the most common environment, and metal-composition engineering as the descriptor most correlated with Faradaic efficiency. On that basis the authors chose CoOx as the platform for modelling and validation.^[S1]

The model: three kinds of input, one regression target

Text fields are encoded by MaterialsBERT and reduced with PCA; literature-reported reaction pathways such as *NO₃→*NO₂ are split into symbol sequences and encoded by ReactionLSTM; metal information enters as numerical features; a DNN regresses Faradaic efficiency on the concatenation. The authors state plainly that the pathways are statistical proxies for how reactions are reported, not taken as physical mechanisms.^[S1]

A random train/test split left the two sets more than 58% similar on average, so information leaked. The authors instead split by catalyst type, holding out a random subset of types that never appear in training. Under that split MatSemNet reaches a training R² of 0.8203 and a test R² of 0.727, with training MAE 0.0225 and test MAE 0.0891.^[S1][S2]

0.727 vs 0.3147

test R² of MatSemNet and of five conventional models trained on numerical and categorical features only (random forest, SVR, XGBoost, GBDT, LightGBM) under the same grouped split; the best of those is random forest at 0.3147^[S2]

+31.56%

relative gain in test R² from adding ReactionLSTM with the MaterialsBERT encoder; 11.61%, 9.72% and 40.22% with SciBERT, MatBERT and Text-embedding-3-large^[S1]

What the model proposed

The authors fixed the CoOx system, added La, Ce, Pr, Nd, Sm, Eu and Gd in turn, and varied electrolyte and voltage with every other feature held constant, asking the model for each one's Faradaic efficiency. In acidic, neutral and alkaline media alike the model put the La-containing system highest. The predicted alkaline order was La > Ce > Pr > Nd > Sm > Gd > Eu. The model's absolute values were highest in acid; the authors caution that hydrogen evolution competes strongly there, so predictions in that regime are read only as relative trends.^[S1]

Rare earths do appear in the training set (La 10 times, Ce 11, Pr 3, Nd 2, Eu 2, Gd 1, Sm none), but by the authors' check none is a simple rare-earth-doped spinel Co₃O₄ matching the experimental series.^[S2]

What people decided

The model did not pick this family from open chemical space. CoOx was chosen by the authors from the literature statistics; the rare-earth series was chosen on chemical grounds: most rare earths sit at a stable +3 state while ionic radius and 4f configuration vary systematically, making them a controlled comparison. The paper also says the compositions 'were not predefined based on prior experimental preference but emerged from model-based screening'. Both sentences stand in the paper; the accurate reading is that the authors bounded the family and the model ordered it.^[S1]

The electrolyte to test was a human call too. The model predicted the highest efficiencies in acid; the authors, weighing dominant hydrogen evolution in acid and slow kinetics in neutral media, made alkaline the focus. Faradaic efficiency and yield were quantified only in 1 M KOH; neutral and acidic media got LSV currents only.^[S1]

The authors themselves position the model as a trend-oriented screening and prioritization tool, not a surrogate for performance: agreement between prediction and experiment is 'mainly reflected at the performance-ranking level'. The numbers say the same. In the paper's Fig. 3d the seven dopants' alkaline predictions are drawn on a radial axis running from 84.0% to 85.4%, while the model's MAE on the held-out test set is 0.0891. Read on a 0-1 scale for Faradaic efficiency (the paper does not state the scale), that is about nine percentage points, far wider than the predicted gaps between dopants. What the seven samples tested is an order, not an absolute efficiency. The reverse holds too: the test R² measures fit to literature data and is not evidence that the ranking is right; only the seven-sample experiment speaks to that.^[S1][S2]

Making and testing all seven is the most valuable choice in the paper's experimental design. Making only the top pick shows at most that the model's choice was decent; making the whole series tests the order itself. The cost is that the sample is seven, quantified in one electrolyte, the bottom three are indistinguishable in the measurements, and the paper does not say how many times each point was repeated.

The ranking test still lacks a control: would a chemist's intuition have made the same first pick? La-doped Co₃O₄ for nitrate electroreduction had already been published (for example a Science China Materials paper on La-doped electron-deficient Co₃O₄ and a Journal of Rare Earths paper on La-doped Co₃O₄ nanoneedle arrays, both 2025), and the paper's own comparison table lists a La-Co₃O₄/CC at 96.4% Faradaic efficiency and 9.14 mg h⁻¹ cm⁻² in 0.1 M NaOH. The model's first pick was right, but it was not a material nobody had thought of.^[S2][S3][S4]

What was measured.

All seven samples are nanorods that keep the spinel structure; the main XRD peaks shift to lower angle and the Raman A1g mode moves, pointing to lattice incorporation. ICP-OES puts rare-earth content between 1.14 and 1.28 wt%, nearly equal across samples (nominal loading 5 mol%). In 1 M KOH the measured order of Faradaic efficiency and yield is La > Ce > Pr > Nd > Sm ≈ Gd ≈ Eu (the order as the paper's text states it); of the seven, only La's values are printed, the rest are drawn as bars in Fig. 5d. La-Co₃O₄ reaches 85.5% and 28.3 mg h⁻¹ cm⁻², which the text places at −0.5 V vs. RHE after 1 h of electrolysis (see below on the potential), and ran 30 one-hour chronoamperometry cycles at −0.6 V in 1.0 M KOH with 0.1 M NO₃⁻. Fig. 5d is labelled 1 M KOH and gives no nitrate concentration.^[S1][S2]

Against which baseline.

The direct performance control is undoped Co₃O₄ in the same lab under the same conditions; La-Co₃O₄ and Ce-Co₃O₄ are also set against literature Co-based catalysts, but those numbers come from other electrolytes, potentials and electrodes. The control for predictive skill is five conventional ML models: MatSemNet has the highest test R², yet its test MAE (0.0891) is the largest of the six (the worst, since lower is better), with the five baselines at 0.0672-0.0766; the paper concedes some tree models reach 'comparable' MAE. These are fit metrics on literature data and say nothing about whether the ranking is right; the evidence for the ranking is the seven-sample experiment alone. No control for selection (a random pick or a chemist's pick) was run.^[S1][S2]

Known limitation: the mechanism fits the order only in part.

DFT nitrate adsorption energies are La −3.41, Ce −2.02, Pr −1.88 and Nd −1.65 eV, with the heavier rare earths and undoped Co₃O₄ between −0.94 and −1.13 eV, matching the measured order, though the bottom three (Sm −0.97, Gd −1.00, Eu −1.13 eV) are bunched too, and in the reverse of the model's predicted order. In situ Raman shows NH₃ signals on La-Co₃O₄ 100 mV earlier than on undoped Co₃O₄. The defect account runs against the measurements: every doped sample has a higher Co²⁺/Co³⁺ ratio than undoped Co₃O₄'s 1.62, which the paper reads as rare earths adding defects, and its Discussion credits light rare earths with inducing surface defects; but the ratio climbs from 1.68 for La to 1.85 for Eu, the reverse of the performance order. On the paper's own reading the best sample, La, has the fewest defects, so 'rare earths add defects' cannot explain the ranking.^[S1][S2]

Inconsistencies and gaps in the preview.

The text places the 85.5% and 28.3 mg h⁻¹ cm⁻² at −0.5 V vs. RHE, while in Fig. 5d the tallest La-Co₃O₄ Faradaic-efficiency bar belongs to the −0.6 V series, the −0.5 V efficiency and yield bars are both clearly lower, and the stability run is at −0.6 V; the calcination temperature is 350 °C in the Results and Fig. 4a but 300 °C in the Methods. The Methods have no electrochemistry section: the cell and the reference electrode are not described, and the electrode area S that the yield equation divides by is never defined; nor is what the error bars show or how many repeats stand behind them. As of 27 September 2026 the accepted preview is still the only version online (no Version of Record yet, and no arXiv or other preprint found), so each of these points should be checked against the Version of Record.^[S1]

A literature model fits ranking within a family

Literature data are heterogeneous and noisy, so a model trained on them gets absolute values wrong; within a chemically comparable series, where the errors are roughly shared, the order can still be useful. This paper used the model where it is strongest and described it honestly as such.

'Model-designed catalyst' should be read narrowly

The authors chose the family, the element range and the test electrolyte; the model supplied an order within them; the lab made and tested every candidate. The top pick, La-Co₃O₄, had been reported before. The accurate statement is that a literature-trained model correctly ordered a dopant series the authors chose, not that a model discovered a new catalyst.^[S1][S3]

A higher R² is not a smaller error

MatSemNet's test R² is far above the conventional baselines while its test MAE is the largest: the worst point-by-point error of the six. It reproduces the test set's variance better without being more accurate point by point, consistent with using it to screen and rank candidates rather than to predict absolute efficiency, and a reason to read both metrics in papers like this. Keep the two layers of evidence apart as well: the paper cites R² 0.727 as support for reliable ranking, but no fit metric proves a ranking; the ranking conclusion here rests only on the seven-sample experiment.^[S2]

What would change the judgment

If the model's order held beyond Co-based oxides on candidates registered in advance, or its first pick beat one a chemist made independently, the work would move from 'ranked a known family correctly' to 'can steer a search'. The dataset and code are available only on request from the corresponding authors; independent reproduction starts when they are released.^[S1]