Source-linked AI summary
A literature-guided descriptor-based framework for filtering composition search spaces
Lei Zhang, Markus Stricker
TL;DR
Large composition search spaces are difficult to screen, and fixed descriptors may not suit changing performance metrics. This paper selects metric-dependent descriptor pairs from literature-trained embeddings and uses them in a Pareto-based composition filter. Across the evaluated metrics, the approach reduces candidate spaces while preserving best or near-best compositions, with a more controlled balance than expert-chosen or random descriptors.
Problem
Large composition spaces can exceed experimental screening capacity, while fixed descriptors may be unsuitable when the performance metric changes.
Source: §I · pp. 1–2Method
For each performance metric, the framework systematically selects two descriptors from a filtered literature-trained embedding vocabulary and uses them to construct a Pareto-based composition filter.
Source: A. Selecting descriptors in embedding… · p. 2, §III · p. 5Results
Current density retains 6.87 ± 2.36% of candidates with 3.85 ± 6.38% average best-value error, while onset overpotential, resistance, and stability have 0.00 ± 0.00% average error in evaluated results.
Source: §III · p. 6Takeaways & Limitations
Framework-selected descriptors vary with performance metric and provide a more controlled balance between retained fraction and best-value error than expert-chosen or random descriptors.
Source: §III · p. 6Onset overpotential and resistance are each evaluated on one search space, and the three stability definitions share one Fe-Co-Ni space; broader validation across datasets, metrics, corpora, and embedding constructions is needed.
Source: §III · p. 6
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
Composition searches can exceed experimental capacity, while literature embeddings offer contextual material knowledge that can guide reduction. The framework addresses fixed-descriptor limitations by selecting metric-dependent descriptor pairs and using them in a Pareto filter.
I. INTRODUCTION
- Large composition spaces are difficult to screen exhaustively, especially for multicomponent systems, so filtering must narrow candidates while retaining high performers.Only a limited number of candidates can be examined in detail experimentally or through more expensive computation.Source: §I · p. 1
- Fixed expert-chosen descriptors may miss useful embedding information, remain heuristic, and be unsuitable when the performance metric changes.The desired descriptors should remain inspectable, comparable across metrics, and traceable to the literature-derived embedding.Source: §I · pp. 1–2
- For each performance metric, the framework selects two descriptors from a literature-trained embedding and retains non-dominated compositions in their two-dimensional similarity space.The two-descriptor representation permits a trade-off-based Pareto filter while keeping the projection interpretable.Source: §I · p. 2
A. Selecting descriptors in embedding space
- Descriptor pairs balance lower and higher similarity to the metric with similarity between descriptors, then the final Pareto-front pair is chosen by average similarity to “performance.”For current density, cyclability and rhombic lie on the descriptor-level Pareto front and combine metric association with contrast.Source: A. Selecting descriptors in embedding… · p. 2
- The selected descriptors occupy the same local embedding region as current density and performance, with cyclability closer to both and rhombic more separated.Selection and similarity calculations use the 200-dimensional embedding; the two-dimensional PCA view is for visualization.Source: A. Selecting descriptors in embedding… · p. 2
B. Composition filtering for current density
For current density, the framework substantially reduces candidate sets while retaining strong compositions. Across other metrics, retained fractions are broader but the evaluated subsets preserve the best-performing compositions.
B. Composition filtering for current density
- 6.87±2.36% of current-density candidates are retained on average across five search spaces, with 3.85 ± 6.38% average best-value error.The retained subset includes the full-space best composition in three of the five search spaces.Source: B. Composition filtering for current… · p. 2
- Ag-Au-Pd-Pt-Rh falls from 327 to 24 candidates while retaining the best composition; Ag-Pd-Pt has the least favorable result, with 14 of 341 retained and 16.41% error.Ag-Au-Pd-Pt-Ru falls from 335 to 14 candidates with 2.84% error, while Ag-Pd-Pt-Ru and Ag-Pd-Ru retain the best composition.Source: B. Composition filtering for current… · p. 3
C. Generalization to other performance metrics
- Onset overpotential retains 10 of 22 Fe-Co-Ni candidates, resistance retains 94 of 342 Ir-Pd-Pt-Rh-Ru candidates, and each stability definition retains 11 of 22 Fe-Co-Ni candidates.Each reported case retains its best-performing composition.Source: C. Generalization to other performance… · p. 3
- For onset overpotential, resistance, and stability, the evaluated filtering results preserve the best-performing compositions with zero best-value error, while retaining broader subsets than current density.The stability results cover three time-interval definitions from the same Fe-Co-Ni search space.Source: C. Generalization to other performance… · pp. 3–4
D. Baseline comparisons
The framework-selected descriptors are compared with a fixed expert-chosen pair and random descriptor pairs. The comparisons show that stronger reduction or occasional best-composition retention does not ensure the same balance of retained fraction and error.
D. Baseline comparisons
- For current density, expert descriptors retain 4.21 ± 1.04% versus 6.87 ± 2.36% for framework descriptors, but increase average best-value error from 3.85 ± 6.38% to 6.01 ± 8.83%.For resistance, the expert pair retains 8.77 ± 0.00% with 1.67±0.00% error, whereas framework descriptors retain 27.49 ± 0.00% with the lowest-resistance composition retained exactly.Source: D. Baseline comparisons · p. 4
- Random descriptors do not consistently match the framework’s balance: for current density they retain 22.10 ± 8.54% with 5.44 ± 3.30% error, versus 6.87 ± 2.36% and 3.85 ± 6.38%.For onset overpotential, the framework retains 45.45 ± 0.00% versus 57.00 ± 16.51% for random descriptors.Source: D. Baseline comparisons · p. 4
- For stability, framework descriptors yield zero best-value error across all three intervals, whereas random descriptors average 218.02±257.04% error and reach 848.22% in the worst seed.For resistance, framework descriptors retain 27.49±0.00% with 0.00 ± 0.00% error, compared with 15.28 ± 10.71% and 2.57 ± 4.60% for random descriptors.Source: D. Baseline comparisons · p. 4
III. DISCUSSION
The framework uses metric-dependent descriptors to reduce composition spaces while retaining the best or near-best compositions, with a more controlled balance between retained fraction and best-value error than fixed or random descriptors. Descriptor selection balances association with the performance metric and contrast between descriptors, but the study’s scope and vocabulary filtering constrain its conclusions.
III. DISCUSSION
- Descriptor selection balances similarity to the performance metric with contrast between descriptors, favoring complementary rather than redundant filtering directions.The ranking also considers similarity to the word performance; the authors note that embedding associations lack a formal explanation for downstream usefulness [14].Source: §III · pp. 5–6
- Selected descriptors vary with performance metric but can recur: cyclability and rhombic appear for current density and stability, while other metrics yield different pairs.Onset overpotential selects instability and spce; capacity and intermetallic are selected for resistance.Source: §III · p. 6
- Each candidate composition is projected onto two selected descriptors, and non-dominated points in that two-dimensional similarity space define the retained subset.The framework uses Pareto filtering both to select descriptors in embedding space and to retain compositions in descriptor-similarity space.Source: §III · p. 6
- Current density retains 6.87 ± 2.36% of candidates with 3.85 ± 6.38% average best-value error, while onset overpotential, resistance, and stability show 0.00 ± 0.00% error.Retained fractions are larger for onset overpotential, resistance, and stability; the useful outcome is reduced search space while preserving the optimum or a near-optimum composition.Source: §III · p. 6
- Framework-selected descriptors provide a more controlled balance between retained fraction and best-value error than expert-chosen or random descriptors.Fixed dielectric and conductivity descriptors can yield stronger reduction with larger best-value error, while random-descriptor outcomes vary across seeds and metrics, especially for stability.Source: §III · p. 6
- Validation is limited across metrics and search spaces, and corpus-specific or ambiguous descriptors such as spce show that vocabulary filtering needs further attention.Current density covers five composition spaces; onset overpotential and resistance each cover one, while three stability definitions use the same Fe-Co-Ni space. The study also does not establish that its suffix patterns or two-descriptor construction are uniquely optimal.Source: §III · pp. 5–7
IV. CONCLUSION
The framework selects two descriptors from a prefiltered, literature-trained embedding vocabulary for each performance metric and uses them in a Pareto-based composition filter. It reduces evaluated composition spaces while preserving the best or near-best measured compositions, with a controlled balance between retained fraction and best-value error compared with expert-chosen or random descriptors.
IV. CONCLUSION
- Across evaluated metrics, the framework reduces candidate composition spaces while preserving the best or near-best measured compositions.Its performance metric-dependent descriptors provide a reproducible, controlled balance between retained fraction and best-value error compared with expert-chosen and random descriptors.Source: §IV · p. 7
- For each performance metric, the framework selects two descriptors from a prefiltered vocabulary in a literature-trained word embedding space to construct a Pareto-based filter.Cyclability and rhombic recur for current density and stability.Source: §IV · p. 7
- Literature-derived embeddings support interpretable and reproducible filters for narrowing candidate compositions.The selected descriptors vary with performance metric, while providing a controlled balance against expert-chosen and random descriptors.Source: §IV · p. 7
V. METHODS
The framework represents performance metrics and candidate compositions in a literature-trained embedding space, then uses selected descriptors to Pareto-filter compositions. Metric phrases are averaged from in-vocabulary words, and candidate descriptors come from a suffix-filtered vocabulary.
V. METHODS
- The two-part framework selects descriptors from a filtered vocabulary for a performance metric, projects compositions onto them, and applies Pareto filtering to produce a retained subset and evaluation measures.The output includes the retained fraction, best-value error, and whether the full-space best-performing composition is retained.Source: §V · p. 7
A. Word embeddings and performance metric representation
- A multi-word performance metric outside the vocabulary is split into component words, and its vector is the normalized average of the in-vocabulary components.This lets single-word and multi-word metrics share the descriptors' embedding space.Source: A. Word embeddings and performance… · p. 7
B. Descriptor vocabulary
- Candidate descriptors are vocabulary words meeting length, alphabetic-character, lowercase, and suffix-pattern requirements.The listed suffix patterns define the filtered descriptor vocabulary.Source: B. Descriptor vocabulary · p. 7
C. Descriptor selection
The method evaluates all unordered pairs of candidate descriptors and uses a three-objective Pareto selection to identify a compromise pair for each performance metric. It then selects one pair from the descriptor-level Pareto front by average similarity to the word performance.
C. Descriptor selection
- All unordered two-descriptor combinations are evaluated; using two descriptors is an exploratory choice for an interpretable trade-off filter, not a claim of unique optimality.The two descriptors define a two-dimensional projection.Source: C. Descriptor selection · p. 8
- Three-objective Pareto selection minimizes the lower descriptor–metric similarity and pairwise descriptor similarity while maximizing the higher descriptor–metric similarity.This tests the hypothesis that one descriptor stays relatively close to the metric while the other provides contrast, with redundant pairs discouraged.Source: C. Descriptor selection · p. 8
- The descriptor combination with the largest average similarity to the word performance is selected from the descriptor-level Pareto front.The front may contain multiple combinations.Source: C. Descriptor selection · p. 8
D. Composition representation and descriptor-similarity projection
Candidate compositions are represented by normalized vectors generated from their elemental fractions and projected into a two-dimensional space using similarities to the selected descriptors. The retained subset is the union of two Pareto fronts that favor opposite directions across the descriptors.
D. Composition representation and descriptor-similarity projection
- Each candidate composition is projected into a two-dimensional descriptor-similarity space, with one similarity value for each selected descriptor.Each composition is represented as a point in that projection space.Source: D. Composition representation and… · p. 8
- The retained subset is the union of two Pareto fronts: one maximizes similarity to the first descriptor while minimizing similarity to the second, and the other reverses those directions.Composition vectors are generated from elemental fractions and normalized; measured performance values are used only after filtering for evaluation.Source: D. Composition representation and… · pp. 8–9
E. Evaluation metrics
Evaluation records how much of the composition search space remains and how well the retained subset preserves the optimum. It also tracks whether the full-space best composition is retained and the rank of the best retained composition.
E. Evaluation metrics
- The retained fraction measures the share of the full composition search space kept; lower values indicate stronger reduction.The full space contains N candidates.Source: E. Evaluation metrics · p. 9
- Best-value error compares the best value in the retained subset with the full-space best, using the metric's direction for minimization or maximization.The method also records whether the full-space best-performing composition is retained and the rank of the best retained composition in the full space.Source: E. Evaluation metrics · p. 9
F. Baseline settings
The baselines bypass descriptor selection, applying expert-chosen descriptors or randomly sampled descriptor pairs to the same composition-filtering procedure. This comparison isolates the effect of performance metric-dependent descriptor selection.
F. Baseline settings
- The expert-chosen baseline applies dielectric and conductivity to every evaluated performance metric.The descriptor-selection step is bypassed.Source: F. Baseline settings · p. 9
- The random baseline samples two descriptors per seed from a shared vocabulary and evaluates seeds 0 through 99.The shared vocabulary is the sorted intersection of suffix-filtered descriptor vocabularies from the evaluated performance metrics.Source: F. Baseline settings · p. 9
- Both baselines retain the same composition representation, similarity projection, bidirectional two-objective Pareto filter, and evaluation metrics.Keeping the filtering procedure unchanged isolates the effect of descriptor selection from fixing descriptors or sampling them randomly.Source: F. Baseline settings · p. 9
Appendix A: Supplementary Information
The supplementary tables provide detailed filtering results across the evaluated performance metrics and descriptor baselines, while the supplementary figures show metric-specific descriptor selections.
SUPPLEMENTARY TABLES
- The detailed tables report filtering counts, best values, and best-value errors for current density, onset overpotential, and resistance.Current-density tables specify the applied potential; onset-overpotential results are in mV, and resistance tables report lowest-resistance values.Source: Table A1 · p. 11, Table A2 · p. 11, Table A3 · p. 11
- Stability tables report percentage-change results, with smaller values indicating higher stability; the random-descriptor summary notes that unsuitable pairs can yield very large errors.The stability cases include different time intervals, and the random assessment summarizes results across composition search spaces over 100 seeds.Source: Table A4 · p. 11, Table A9 · p. 12
- Expert-descriptor baseline tables report current density, onset overpotential, resistance, and stability filtering results using dielectric + conductivity.The stability table includes three time-interval definitions from the same Fe-Co-Ni composition search space.Source: Table A5 · p. 12, Table A6 · p. 12, Table A7 · p. 12
SUPPLEMENTARY FIGURES
- The supplementary figures show selected descriptor pairs for onset overpotential (instability and spce), resistance (capacity and intermetallic), and stability (cyclability and rhombic).Each figure also depicts descriptor-level Pareto selection and a PCA projection of the filtered embedding space.Source: Figure A1 · p. 13, Figure A2 · p. 13, Figure A3 · p. 14