A literature-guided descriptor-based framework for filtering composition search spaces

Lei Zhang; Markus Stricker

cs.CLSep 14, 2026

A literature-guided descriptor-based framework for filtering composition search spaces

Authors: Lei Zhang, Markus Stricker

Organizations: Interdisciplinary Centre for Advanced Materials Simulation, Ruhr University Bochum, Universitätsstraße 150, 44780 Bochum, North Rhine-Westphalia, Germany

Abstract

Scientific literature contains latent knowledge about materials behavior, but much of this knowledge is expressed through words, contexts, and recurring associations rather than explicit design principles. This raises a central question: how can large-scale scientific corpora be used for practical problems in materials discovery? Here, we present a literature-guided descriptor-based filtering framework for reducing composition search spaces. For a given performance metric, the framework selects two descriptors from a filtered vocabulary in a literature-trained word embedding model and uses the selected descriptors to construct a Pareto-based filter for candidate compositions. Across the evaluated performance metrics and composition search spaces, the framework filters out an average of 74.27% of the candidate compositions, with an average best-value error of 1.93% relative to experimental measurements. Compared with expert-chosen and random descriptors, our performance metric-dependent descriptors provide a more controlled balance between retained fraction and best-value error. These results show that literature-derived embeddings can support intuitive and reproducible filters for narrowing candidate composition spaces while preserving high-performing compositions.

Figures & tables

Figure 1: Descriptor selection for the performance metric current density . (a) Possible two-descriptor combinations in the three-objective selection space. The Pareto-based selection minimizes the lower similarity to current density , maximizes the higher similarity to current density , and minimizes the cosine similarity between the two descriptors. Open markers indicate the descriptor-level Pareto front, and the selected descriptors are cyclability and rhombic . Similarity to the word performance is not an axis in (a); it is used only in the second stage to select the final descriptor combination from the Pareto front. (b) Two-dimensional PCA projection of the 200-dimensional filtered embedding space, showing current density , the word performance , and the two selected descriptors.

Performance metric	Selected descriptors	Avg. ret.	Avg. err.
Current density	cyclability + rhombic	$6.87\pm 2.36$	$3.85\pm 6.38$
Onset overpotential	instability + spce	$45.45\pm 0.00$	$0.00\pm 0.00$
Resistance	capacity + intermetallic	$27.49\pm 0.00$	$0.00\pm 0.00$
Stability	cyclability + rhombic	$50.00\pm 0.00$	$0.00\pm 0.00$

Table 1: Summary of the literature-guided descriptor-based filtering framework. Selected descriptors gives the two descriptors selected for each performance metric. Avg. ret. is the average retained fraction in percent. Avg. err. is the average best-value error in percent relative to the full composition search space. Values are reported as mean

\pm

population standard deviation across the corresponding filtering results. For stability, the averages are calculated from three stability definitions based on different time intervals in the same Fe-Co-Ni composition search space.

Figure 2: Retained-fraction and best-value-error summary. (a) Relation between retained fraction and best-value error, with points colored by performance metric. Each point corresponds to one filtering result. (b) Average retained fraction for each performance metric. (c) Average best-value error for each performance metric on a 0–100% scale. Error bars in (b) and (c) show population standard deviations across the corresponding filtering results.

Performance metric	Expert-chosen descriptors	Avg. ret.	Avg. err.
Current density	dielectric + conductivity	$4.21\pm 1.04$	$6.01\pm 8.83$
Onset overpotential	dielectric + conductivity	$63.64\pm 0.00$	$0.00\pm 0.00$
Resistance	dielectric + conductivity	$8.77\pm 0.00$	$1.67\pm 0.00$
Stability	dielectric + conductivity	$63.64\pm 0.00$	$0.00\pm 0.00$

Table 2: Summary of the expert-chosen descriptor baseline. The baseline uses dielectric and conductivity for all evaluated performance metrics. Values are reported as mean

\pm

population standard deviation across the corresponding filtering results. Avg. ret. is the average retained fraction in percent, and Avg. err. is the average best-value error in percent relative to the full composition search space. For stability, the averages are calculated from three stability definitions based on different time intervals in the same Fe-Co-Ni composition search space.

Figure 3: Schematic of the literature-guided descriptor-based filtering framework.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Composition search space	Applied potential	Selected descriptors	Full count	Retained count	Best/full	Best/retained	Best-value error (%)
Ag-Au-Pd-Pt-Rh	-300 mV	cyclability + rhombic	327	24	-1.13	-1.13	0.00
Ag-Au-Pd-Pt-Ru	-300 mV	cyclability + rhombic	335	14	-1.49	-1.45	2.84
Ag-Pd-Pt	850 mV	cyclability + rhombic	341	14	-0.58	-0.49	16.41
Ag-Pd-Pt-Ru	850 mV	cyclability + rhombic	341	31	-0.37	-0.37	0.00
Ag-Pd-Ru	850 mV	cyclability + rhombic	342	33	-0.67	-0.67	0.00

Appendix

Table A1: Detailed composition-filtering results for the current density performance metric. The applied potential identifies the current-density measurement used for each composition search space. Full count and retained count give the number of candidate compositions before and after Pareto filtering. Best/full and best/retained report the best current-density values in mA cm -2 , and best-value error is reported relative to the full composition search space.

Composition search space	Selected descriptors	Full count	Retained count	Best/full	Best/retained	Best-value error (%)
Fe-Co-Ni	instability + spce	22	10	394.00	394.00	0.00

Appendix

Table A2: Detailed composition-filtering result for the onset overpotential performance metric. Full count and retained count give the number of candidate compositions before and after Pareto filtering. Best/full and best/retained are reported in mV, and best-value error is reported relative to the full composition search space.

Composition search space	Selected descriptors	Full count	Retained count	Best/full	Best/retained	Best-value error (%)
Ir-Pd-Pt-Rh-Ru	capacity + intermetallic	342	94	1.49	1.49	0.00

Appendix

Table A3: Detailed composition-filtering result for the resistance performance metric. Full count and retained count give the number of candidate compositions before and after Pareto filtering. Best/full and best/retained report the lowest-resistance values, and best-value error is reported relative to the full composition search space.

Composition search space	Time interval	Selected descriptors	Full count	Retained count	Best/full	Best/retained
Fe-Co-Ni	$e_{0}\to e_{1}$	cyclability + rhombic	22	11	0.59	0.59
Fe-Co-Ni	$e_{1}\to e_{2}$	cyclability + rhombic	22	11	0.20	0.20
Fe-Co-Ni	$e_{0}\to e_{2}$	cyclability + rhombic	22	11	0.40	0.40

Appendix

Table A4: Detailed composition-filtering results for the stability performance metric. All three rows come from the same Fe-Co-Ni composition search space but use different stability definitions based on different time intervals. Here

e_{0}

is the onset overpotential in mV,

e_{1}

is the overpotential after 1 h, and

e_{2}

is the overpotential after 2 h. The time interval gives the percentage-change interval used to define stability. Full count and retained count give the number of candidate compositions before and after Pareto filtering. Best/full and best/retained are reported as percentage change, where smaller values indicate higher stability.

Composition search space	Applied potential	Expert-chosen descriptors	Full count	Retained count	Best/full	Best/retained	Best-value error (%)
Ag-Au-Pd-Pt-Rh	-300 mV	dielectric + conductivity	327	9	-1.13	-1.09	3.80
Ag-Au-Pd-Pt-Ru	-300 mV	dielectric + conductivity	335	20	-1.49	-1.45	2.84
Ag-Pd-Pt	850 mV	dielectric + conductivity	341	15	-0.58	-0.45	23.40
Ag-Pd-Pt-Ru	850 mV	dielectric + conductivity	341	13	-0.37	-0.37	0.00
Ag-Pd-Ru	850 mV	dielectric + conductivity	342	14	-0.67	-0.67	0.00

Appendix

Table A5: Detailed composition-filtering results for the current density performance metric using the expert-chosen descriptor baseline dielectric + conductivity . The applied potential identifies the current-density measurement used for each composition search space. Full count and retained count give the number of candidate compositions before and after Pareto filtering. Best/full and best/retained report the best current-density values in mA cm -2 , and best-value error is reported relative to the full composition search space.

Composition search space	Expert-chosen descriptors	Full count	Retained count	Best/full	Best/retained	Best-value error (%)
Fe-Co-Ni	dielectric + conductivity	22	14	394.00	394.00	0.00

Appendix

Table A6: Detailed composition-filtering result for the onset overpotential performance metric using the expert-chosen descriptor baseline dielectric + conductivity . Full count and retained count give the number of candidate compositions before and after Pareto filtering. Best/full and best/retained are reported in mV, and best-value error is reported relative to the full composition search space.

Composition search space	Expert-chosen descriptors	Full count	Retained count	Best/full	Best/retained	Best-value error (%)
Ir-Pd-Pt-Rh-Ru	dielectric + conductivity	342	30	1.49	1.47	1.67

Appendix

Table A7: Detailed composition-filtering result for the resistance performance metric using the expert-chosen descriptor baseline dielectric + conductivity . Full count and retained count give the number of candidate compositions before and after Pareto filtering. Best/full and best/retained report the lowest-resistance values, and best-value error is reported relative to the full composition search space.

Composition search space	Time interval	Expert-chosen descriptors	Full count	Retained count	Best/full	Best/retained
Fe-Co-Ni	$e_{0}\to e_{1}$	dielectric + conductivity	22	14	0.59	0.59
Fe-Co-Ni	$e_{1}\to e_{2}$	dielectric + conductivity	22	14	0.20	0.20
Fe-Co-Ni	$e_{0}\to e_{2}$	dielectric + conductivity	22	14	0.40	0.40

Appendix

Table A8: Detailed composition-filtering results for the stability performance metric using the expert-chosen descriptor baseline dielectric + conductivity . All three rows come from the same Fe-Co-Ni composition search space but use different stability definitions based on different time intervals. Here

e_{0}

is the onset overpotential in mV,

e_{1}

is the overpotential after 1 h, and

e_{2}

Performance metric	Seeds	Mean ret. (%)	SD ret. (%)	Mean err. (%)	SD err. (%)	Best err. (%)	Worst err. (%)
Current density	100	22.10	8.54	5.44	3.30	0.09	19.18
Onset overpotential	100	57.00	16.51	0.76	3.79	0.00	22.08
Resistance	100	15.28	10.71	2.57	4.60	0.00	20.24
Stability	100	57.00	16.51	218.02	257.04	0.00	848.22

Appendix

Table A9: Summary of the 100-seed random-descriptor assessment. For each seed, two descriptors were sampled from the shared descriptor vocabulary and then applied to all evaluated filtering results. Reported values are performance-metric-level summaries across the corresponding composition search spaces. Standard deviations are population standard deviations across the 100 seeds. Because the stability metric is a percentage change with best values close to zero, unsuitable random descriptors can produce very large best-value errors for that metric.

Figure A1: Descriptor selection for the performance metric onset overpotential . (a) Possible two-descriptor combinations in the three-objective selection space and the descriptor-level Pareto front. (b) PCA projection of the filtered embedding space, showing onset overpotential , the word performance , and the selected descriptors instability and spce .

Figure A2: Descriptor selection for the performance metric resistance . (a) Possible two-descriptor combinations in the three-objective selection space and the descriptor-level Pareto front. (b) PCA projection of the filtered embedding space, showing resistance , the word performance , and the selected descriptors capacity and intermetallic .

Figure A3: Descriptor selection for the performance metric stability . (a) Possible two-descriptor combinations in the three-objective selection space and the descriptor-level Pareto front. (b) PCA projection of the filtered embedding space, showing stability , the word performance , and the selected descriptors cyclability and rhombic .

Materials Discovery

See all 29 papers

From Word2Vec to Transformers: Text-Derived Composition Embeddings for Filtering Combinatorial Electrocatalysts
Mar 9, 2026Lei Zhang, Markus StrickerVirtual Screening Compositional Representation Learning
Interdisciplinary Centre for Advanced Materials Simulation, Ruhr University Bochum, Universitätsstraße 150, 44801 Bochum, Germany
Agentic Design of Compositional Descriptors via Autoresearch for Materials Science Applications
May 14, 2026Matteo Cobelli, Stefano SanvitoMaterials Property Prediction Automated Feature Engineering
School of Physics and CRANN Institute, Trinity College, Dublin 2, Dublin, Ireland.
From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals
May 28, 2026Yeyong Yu, Wenya Hu, Xing Wu +1LLM Evaluation Pairwise Preference Evaluation
School of Computer Engineering & Science, Shanghai University, Shanghai, 200444, China. · Center of Materials Informatics and Data Science, Materials Genome2026 Institute, Shanghai University, Shanghai, 200444, China. · Key Laboratory of Silicate Cultural Relics Conservation (Shanghai University), Ministry of Education, China. +1

Mar 9, 2026J/K move · Enter open · S save

From Word2Vec to Transformers: Text-Derived Composition Embeddings for Filtering Combinatorial Electrocatalysts

Lei Zhang, Markus Stricker

Interdisciplinary Centre for Advanced Materials Simulation, Ruhr University Bochum, Universitätsstraße 150, 44801 Bochum, Germany

Virtual Screening Compositional Representation Learning Text Embedding Evaluation Word Embeddings

Open paper AskPDF arXiv

Compositionally complex solid solution electrocatalysts span vast composition spaces, and even one materials system can contain more candidate compositions than can be measured exhaustively. Here we evaluate a label-free screening strategy that represents each composition using embeddings derived from scientific texts and prioritizes candidates based on similarity to two property concepts. We compare a corpus-trained Word2Vec baseline with transformer-based embeddings, where compositions are encoded either by linear element-wise mixing or by short composition prompts. Similarities to `concept directions', the terms conductivity and dielectric, define a 2-dimensional descriptor space, and a symmetric Pareto-front selection is used to filter candidate subsets without using electrochemical labels. Performance is assessed on 15 materials libraries including noble metal alloys and multicomponent oxides. In this setting, the lightweight Word2Vec baseline, which uses a simple linear combination of element embeddings, often achieves the highest number of reductions of possible candidate compositions while staying close to the best measured performance.

Abstract

Materials Discovery

Explore similar work

From Word2Vec to Transformers: Text-Derived Composition Embeddings for Filtering Combinatorial Electrocatalysts

Agentic Design of Compositional Descriptors via Autoresearch for Materials Science Applications

From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals