Pith. sign in

REVIEW 4 major objections 4 minor 51 references

UniPolymer claims a unified pipeline that predicts polyimide glass-transition temperatures, generates candidate structures conditioned on a target Tg, and filters out off-target candidates, achieving R²=0.93 and a 73.79% valid-candidate rat

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 10:42 UTC pith:TSA5FUYM

load-bearing objection PITg-Curated is a useful new dataset, but the MD validation is not external — the same COMPASS protocol generated 21% of the training labels and then 'confirmed' the generated candidates. the 4 major comments →

arxiv 2607.29256 v1 pith:TSA5FUYM submitted 2026-07-31 cs.LG

UniPolymer: A Unified Framework for Property Prediction, Structure Recommendation, and Evaluation in Polyimide Design

Junquan Hu , Zhihui Wang , Peng Xu , Xinru Guo , Xintong Li , Kun Lu , Ben Fei This is my paper
classification cs.LG
keywords polyimideglass transition temperatureproperty predictionconditional generationSELFIESstructure recommendationmolecular dynamicsmaterials informatics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that polyimide structures with a desired glass transition temperature can be generated and screened computationally within a single framework. UniPolymer couples a Tg predictor, a SELFIES-based conditional generator, and a candidate evaluator that filters structures whose predicted Tg deviates more than 20°C from the target. On a new dataset of 10,066 polyimide repeating units, the pipeline achieves R²=0.93 for Tg prediction and a 73.79% structural pass rate, with recommended candidates matching molecular-dynamics Tg estimates within about 2°C. If true, this would reduce the number of candidates that must be synthesized and tested.

Core claim

UniPolymer's central claim is that enforcing consistency between what a model predicts and what it generates is sufficient to keep generated polyimide structures on target for glass transition temperature. Concretely, after training a Tg regressor on PITg-Curated, the authors freeze it and use it to score every candidate produced by a target-conditioned SELFIES generator; candidates are ranked by absolute Tg deviation and structural constraints, and only those within 20°C of the target are recommended. The reported outcome is a recommendation pass rate of 73.79% and close agreement (MAE=1.95°C) between predicted and MD-simulated Tg for the top 100 candidates per target.

What carries the argument

The load-bearing component is the frozen property predictor used as a critic. Candidate SELFIES sequences are decoded, re-encoded as SMILES, and fed through the same structure–Tg mapping learned on experimental data, while a continuous-discrete Tg embedding (a continuous standardized value summed with a temperature-bin embedding, then layer-normalized) guides the autoregressive transformer. The absolute deviation between predicted and target Tg, combined with polyimide-specific constraints (imide skeleton, molecular weight, heavy-atom count, stable groups), defines the valid candidate set and the ranking.

Load-bearing premise

The dataset labels—especially the 2,129 Tg values computed by the authors' own molecular dynamics simulations—must be accurate enough to serve as ground truth, since the same MD protocol is later used as the external check on recommended candidates.

What would settle it

Compare, for a set of recommended candidates, the model's predicted Tg and the MD-derived Tg against experimentally measured Tg values from actual synthesis; if the experimental MAE greatly exceeds 17.87°C, the claimed reliability fails. Alternatively, check whether experimental Tg values for PITg-Curated structures that also have MD labels differ systematically from the MD labels.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Generation and evaluation become a closed loop, so candidates that miss the target Tg are removed before wet-lab testing.
  • The PITg-Curated dataset offers a public resource for training and benchmarking polyimide Tg models.
  • The same architecture could be transferred to other property-conditioned polymer design tasks, such as dielectric constant or mechanical strength.
  • Using a frozen predictor as a critic makes the evaluation criterion stable across generation runs, enabling fair comparisons across candidate sets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's validation of recommended candidates uses the same MD protocol that produced 21% of the training labels; if COMPASS-force-field Tg values are systematically biased, both the predictor and the MD 'ground truth' share that bias, and the MAE=1.95°C may reflect self-consistency rather than experimental accuracy.
  • Because the pass rate is computed with the same frozen predictor used for screening, the 73.79% figure measures internal agreement between generator and predictor, not agreement with experimental Tg; the real test is synthesis.
  • A natural extension is to replace the hard 20°C threshold with a probabilistic or cost-aware ranking, and to add synthetic-accessibility or uncertainty estimates as post-generation filters.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes UniPolymer, a unified framework that combines Tg property prediction, target-conditioned SELFIES generation, candidate evaluation, and structure recommendation for polyimide design. It also introduces PITg-Curated, a dataset of 10,066 deduplicated polyimide repeating units with Tg labels, of which 2,129 labels (21.15%) were produced by the authors' own molecular dynamics simulations using the COMPASS force field. The method uses masked-language-model pretraining, contrastive structure-consistency learning, multi-scale feature fusion, and a continuous-discrete Tg condition to guide autoregressive generation. Candidates are filtered and ranked by a frozen property predictor plus polyimide-specific structural constraints. The paper reports R²=0.93, MAE=17.87°C for Tg prediction, a structural pass rate of 73.79%, and MAE=1.95°C / RMSE=2.40°C between model predictions and MD estimates for the top recommended candidates.

Significance. If the claims hold, UniPolymer would be a practically useful front-end for Tg-targeted polyimide design, and the PITg-Curated dataset with released code is a potentially valuable community resource. The framework integrates several sensible components—pretraining, contrastive learning, multi-scale descriptors, and continuous-discrete conditioning—and the ablations give some evidence that each component contributes. However, the current evidence is weakened by a load-bearing circularity: target-consistency metrics are computed with the same frozen predictor used for ranking, and the MD validation shares the same force-field/protocol used to generate 21% of the dataset labels. These issues are addressable with source-stratified splits, experimental-only evaluation, independent validation, and uncertainty quantification, but they need to be fixed before the central claims can be accepted.

major comments (4)
  1. [§S2, §S3, and §Experiments] The external validation is not fully external. PITg-Curated contains 2,129 MD-derived Tg labels (21.15%) produced with the COMPASS force field (S2), and the later 'validation' of recommended candidates uses the same simulation protocol (S3, Fig. 4c). If COMPASS-based Tg estimates have a systematic offset from experimental Tg, the model can learn a structure-to-MD-Tg mapping rather than a structure-to-experimental-Tg mapping. The reported R²=0.93 and MD MAE=1.95°C then demonstrate internal consistency with one simulation protocol, not the claimed practical ability to reduce invalid experiments. The paper should report (i) a source-stratified split, (ii) test-set performance on experimentally labeled data only, and (iii) validation against an independent protocol or experimental Tg values.
  2. [§Recommendation and Evaluation, Eqs. (14)–(15); §Evaluation of Target-Tg Conditional Response] The target-consistency metrics are self-referential. Hit@20, Target MAE, and Valid Hit@20 are computed by comparing generated candidates with the frozen predictor's own estimates, while the same predictor is used to rank and select the top candidates. This measures rank-consistency, not agreement with true Tg. The MD comparison in Fig. 4(c) is further weakened by selection bias: only the top-100 candidates per target are submitted to MD, and those candidates are chosen using the very predictor being validated. The manuscript should evaluate on a random or structurally diverse sample, and should report the MD agreement separately from the model's own predicted ranking.
  3. [Tables 1–4] No uncertainty quantification is provided. All headline comparisons are point estimates: R²=0.93 vs. 0.91 for the strongest baseline, and pass rate 73.79% vs. 72.58%. Without multiple random seeds, bootstrap confidence intervals, or repeated dataset splits, it is impossible to tell whether these differences are statistically meaningful. This is especially important because the dataset is constructed from heterogeneous sources (literature, databases, and the authors' own MD), and the random 8:1:1 split may place MD-derived and experimental labels unevenly across folds. Please add error bars or confidence intervals and report the variance across seeds.
  4. [Table 2 vs. Table 4; §Datasets and Settings] The headline 'pass rate' is ambiguous. Table 2 reports Valid = 73.79% as an overall structural validity rate, but Table 4 reports Valid Hit@20 = 50.84% for the same model under the definition 'structural constraints and |T_pred − T*| ≤ 20°C'. The abstract's 'candidate structure evaluation pass rate of 73.79%' appears to refer only to structural validity, not target-property consistency. These two notions should be clearly separated and both reported in the abstract and conclusions, so readers do not conflate structural validity with target-Tg agreement.
minor comments (4)
  1. [Table 1] The baseline row 'polyBARTguoxinru1120@gmail.com' appears to contain a stray email address; this looks like a typesetting error and should be corrected.
  2. [§Conditional SELFIES Autoregressive Generation] The sentence 'the polyimide structure is first converted into SELFIES, the polyimide structure is first converted into SELFIES' is duplicated. Please remove the repetition.
  3. [Fig. 4(c)] The MD comparison should report per-target breakdowns or error bars, not only the pooled average MAE/RMSE. It would also be useful to state how many MD simulations were successfully converged and discarded.
  4. [§Datasets and Settings] The evaluation section mentions the top-100 candidates per target for MD but does not specify how ties are broken or whether the reported 1.95°C MAE is averaged over all four targets equally. Please clarify the aggregation.

Circularity Check

2 steps flagged

Target-consistency pass rate is defined through the frozen predictor used for ranking, and the MD 'validation' reuses the same COMPASS protocol that generated 21% of the training labels.

specific steps
  1. self definitional [Section 4.1 (Datasets and Settings); Section 3.4 (Recommendation and Evaluation), Eq. (12)–(15)]
    "Target-property consistency is evaluated using the frozen Tg predictor. ... The proportion of candidates that simultaneously satisfy both structural constraints and | bT−T ∗| ≤20◦C was defined as the final recommendation pass rate (Valid Hit@20)."

    The 'valid' target-consistency metric is defined on \hat T, the output of the frozen predictor, and the same \hat T is the selection/ranking criterion: Eq. (14) admits candidates only when dj=|\hat Tj−T*|≤δT and Eq. (15) ranks by −dj. Thus Valid Hit@20 (and the reported 50.84%) reports how often the predictor's own scores satisfy a threshold on the same predictor's scores after the candidates were generated under the target condition and then filtered/ranked with that predictor. No independent experimental Tg enters this metric, so the target-consistency pass rate is, by construction, an internal-consistency statistic rather than a validation of candidate Tg.

  2. other [Supplementary S2/S3; Section 4.4 (Evaluation of Target-Tg Conditional Response...)]
    "Of these, 2129 records (21.15%) originated from molecular dynamics simulations conducted in this paper, while 7,937 records (78.85%) were collected from publicly available polymer databases... Interatomic interactions were described using the COMPASS force field... Furthermore, the top 100 candidates under each target Tg condition were selected for molecular dynamics simulations, and the results were compared with model predictions to verify the reliability of the property assessment and recommendation results."

    The MD protocol used as 'verification' (COMPASS force field, Amorphous Cell/Forcite construction, NVT/NPT equilibration, piecewise density–temperature fit in Eqs. S1–S5) is the same protocol that produced 2,129 of the PITg-Curated training labels. Consequently the reported agreement (MAE=1.95°C, RMSE=2.40°C) can reflect consistency between the learned mapping and one simulation protocol whose bias, if any, is already encoded in 21% of the training labels. The validation is therefore not fully external; it is a shared-protocol check, and the top-100 candidates were additionally selected by the very predictor being validated.

full rationale

UniPolymer's property-prediction results (R²=0.93, MAE=17.87°C) are from a held-out test split and are not by themselves circular; the structural pass rate (73.79%) is also an independent rule-based check. The circularity burden comes from the evaluation of the generation/recommendation pipeline. The paper defines target-property consistency entirely through the frozen predictor's own output (|\hat T−T*|), the same output used to rank candidates. Therefore the reported Hit@20/Valid Hit@20/pass-rate and Target MAE are self-referential by construction: they measure the model against itself. The MD comparison is an attempt at external validation, but the protocol is not external: the same COMPASS-based MD procedure generated 2,129/10,066 training labels and then validated the top-100 candidates, so agreement with MD may only show protocol self-consistency. Selection of the top-100 by the predictor compounds this. The paper itself defers wet-experimental validation to future work. These two issues affect the central claim that UniPolymer reduces invalid experiments, so the score is 6 (partial circularity), not a lower score. No load-bearing self-citation chain is present.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claim rests on the accuracy of the PITg-Curated labels, especially the 21% generated by the paper's own MD pipeline, and on several design choices (20°C threshold, structural filters, Tg-bin count, contrastive temperature, MD protocol) that directly influence the reported pass rates and validation agreement. No new physical entities are introduced; the only new artifact is the dataset itself.

free parameters (6)
  • delta_T = 20°C = 20 °C
    Maximum allowed |predicted Tg - target Tg| for a candidate to be valid; hand-chosen, directly controls pass rates.
  • MW / heavy-atom / unstable-group filters = 150-2000 Da; 10-150 heavy atoms
    Hard-coded structural filters in χmol (Eq. 13); chosen by hand, affect structural pass rate.
  • Number of Tg bins KT = not reported
    Discrete conditioning path Eq. 6; value not reported, affects generation conditioning.
  • Contrastive temperature τ = not reported
    Eq. 3; value not reported, affects representation learning.
  • Label smoothing ε = 0.1
    Eq. 11; chosen.
  • MD protocol parameters = 10 chains, 0.6 g/cm³, 9.5 Å cutoff, 300 ps per temperature
    Choices affect all MD-derived Tg labels (21% of dataset): chain length, initial density, cutoff, simulation time, annealing 300-800 K.
axioms (6)
  • domain assumption COMPASS force field accurately estimates polyimide Tg
    Used for all MD simulations in S3; if this force field is biased for polyimides, 21% of dataset labels and the MD validation are biased.
  • domain assumption Piecewise linear density-temperature fit yields a valid Tg
    Eqs. S1-S5 define Tg as the kink in a two-line fit; not all polymers show a sharp kink, so this could mislabel some structures.
  • domain assumption Tg values from different sources are comparable after median aggregation
    S2 aggregates multiple Tg labels per structure with the median, assuming literature, database, and MD values are exchangeable.
  • domain assumption Random split without source stratification is unbiased
    8:1:1 split in 'Datasets and Settings' does not stratify by experimental vs MD-derived label; MD labels may be easier to predict, inflating R².
  • domain assumption Pretraining on PI1M corpus transfers to polyimide structures
    Methodology transfers a general-polymer encoder to the polyimide domain; no demonstration that the pretraining distribution covers polyimide-relevant chemistry.
  • standard math SELFIES guarantees 100% chemically valid sequences
    The paper relies on SELFIES validity to explain 100% RDKit validity; this is a property of the representation, assuming the implementation is correct.

pith-pipeline@v1.3.0-daily-deepseek · 13586 in / 14792 out tokens · 139699 ms · 2026-08-03T10:42:00.237312+00:00 · methodology

0 comments
read the original abstract

Designing polyimide structures with specific glass transition temperatures (Tg) is highly challenging. Existing methods primarily focus on target-conditioned generation, lacking an assessment of the consistency between the generated structure and the target properties. This leads to low-quality candidates deviating from the design objective entering subsequent processes, increasing invalid experiments and prolonging the development cycle. To address this issue, we propose UniPolymer, a unified framework for property prediction, target-conditioned generation, candidate evaluation, and structure recommendation in polyimide design and a dataset containing 10066 deduplicated polyimide repeating units with Tg tags (PITg-Curated) was constructed. To improve the consistency between generated candidate structures and the target Tg, UniPolymer first establishes a reliable structure-property relationship mapping through self-supervised chemical semantic learning, structural consistency enhancement, and multi-scale information fusion. Subsequently, the model employs a continuous-discrete joint Tg representation to guide the autoregressive generation of SELFIES. The generated candidate structures are further evaluated using a frozen property predictor and polyimide-specific structural constraints, and ranked according to their deviation from the target Tg, thereby preventing structures deviating from the target from entering the subsequent validation stage. Experimental results show that UniPolymer achieved a property prediction accuracy of R^2=0.93 and a candidate structure evaluation pass rate of 73.79%, which are 2% and 1.21% higher than the best baseline, respectively. Meanwhile, the predicted Tg values of the recommended candidates are in high agreement with the results of molecular dynamics simulations, thereby reducing the number of candidates that enter the high-cost experimental stage.

Figures

Figures reproduced from arXiv: 2607.29256 by Ben Fei, Junquan Hu, Kun Lu, Peng Xu, Xinru Guo, Xintong Li, Zhihui Wang.

Figure 1
Figure 1. Figure 1: Overview of UniPolymer. By jointly screening based on property consistency and structural rationality, a [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: UniPolymer’s prediction performance of polyimide Tg on the test set. The predicted values and experimental [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Generated and real polyimide structures are compared in terms of chemical-space distributions and structural [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Target-Tg-driven polyimide structure exploration: (a) predicted Tg distributions; (b) Target MAE and Hit@20; [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 1 linked inside Pith

  1. [1]

    Digital Discovery , volume=

    Polyuniverse: Generation of a Large-Scale Polymer Library Using Rule-Based Polymerization Reactions for Polymer Informatics , author=. Digital Discovery , volume=. 2024 , doi=

    2024

  2. [2]

    Science and Technology of Advanced Materials: Methods , volume=

    NIMS Polymer Database PoLyInfo (II): Machine-Readable Standardization of Polymer Knowledge Expression , author=. Science and Technology of Advanced Materials: Methods , volume=. 2024 , doi=

    2024

  3. [3]

    Polymer Property Predictor and Database , year=

  4. [4]

    RDKit: Open-Source Cheminformatics Software , year=

  5. [5]

    Polymers , volume=

    Prediction and Interpretability Study of the Glass Transition Temperature of Polyimide Based on Machine Learning and Molecular Dynamics Simulations , author=. Polymers , volume=. 2025 , doi=

    2025

  6. [6]

    Journal of Chemical Theory and Computation , volume=

    Rapid, Accurate and Reproducible Prediction of the Glass Transition Temperature Using Ensemble-Based Molecular Dynamics Simulation , author=. Journal of Chemical Theory and Computation , volume=. 2025 , doi=

    2025

  7. [7]

    ACS Applied Polymer Materials , volume=

    Predicting the Glass Transition Temperature of Biopolymers via High-Throughput Molecular Dynamics Simulations and Machine Learning , author=. ACS Applied Polymer Materials , volume=. 2024 , doi=

    2024

  8. [8]

    The Journal of Physical Chemistry B , volume=

    How to Determine Glass Transition Temperature of Polymer Electrolytes from Molecular Dynamics Simulations , author=. The Journal of Physical Chemistry B , volume=. 2024 , doi=

    2024

  9. [9]

    Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on , pages=

    Real-time segmentation of on-line handwritten arabic script , author=. Frontiers in Handwriting Recognition (ICFHR), 2014 14th International Conference on , pages=. 2014 , organization=

    2014

  10. [10]

    Soft Computing and Pattern Recognition (SoCPaR), 2014 6th International Conference of , pages=

    Fast classification of handwritten on-line Arabic characters , author=. Soft Computing and Pattern Recognition (SoCPaR), 2014 6th International Conference of , pages=. 2014 , organization=

    2014

  11. [11]

    Solar Energy , volume=

    High-temperature-resistant and colorless polyimide: Preparations, properties, and applications , author=. Solar Energy , volume=. 2020 , publisher=

    2020

  12. [12]

    Polymer , volume=

    High Heat-Resistant Polyimide Films Containing Quinoxaline Moiety for Flexible Substrate Applications , author=. Polymer , volume=. 2020 , doi=

    2020

  13. [13]

    Nature Reviews Materials , volume=

    Design of functional and sustainable polymers assisted by artificial intelligence , author=. Nature Reviews Materials , volume=. 2024 , publisher=

    2024

  14. [14]

    The Journal of Physical Chemistry B , volume=

    Role of intrinsic factors of polyimides in glass transition temperature: An atomistic investigation , author=. The Journal of Physical Chemistry B , volume=. 2019 , publisher=

    2019

  15. [15]

    npj Computational Materials , volume=

    Unified multimodal multidomain polymer representation for property prediction , author=. npj Computational Materials , volume=. 2025 , publisher=

    2025

  16. [16]

    npj Artificial Intelligence , volume=

    POLYT5: an encoder-decoder foundation chemical language model for generative polymer design , author=. npj Artificial Intelligence , volume=. 2026 , publisher=

    2026

  17. [17]

    ACS Polymers Au , volume=

    Data-driven methods for accelerating polymer design , author=. ACS Polymers Au , volume=. 2021 , publisher=

    2021

  18. [18]

    Polymer Chemistry , volume=

    A review on the application of molecular descriptors and machine learning in polymer design , author=. Polymer Chemistry , volume=. 2023 , publisher=

    2023

  19. [19]

    Journal of Polymer Science , volume=

    Determination of glass transition temperature of polyimides from atomistic molecular dynamics simulations and machine-learning algorithms , author=. Journal of Polymer Science , volume=. 2020 , publisher=

    2020

  20. [20]

    synthetic

    Machine learning with enormous “synthetic” data sets: predicting glass transition temperature of polyimides using graph convolutional neural networks , author=. ACS omega , volume=. 2022 , publisher=

    2022

  21. [21]

    Macromolecules , volume=

    Heat-Resistant Polymer Discovery by Utilizing Interpretable Graph Neural Network with Small Data , author=. Macromolecules , volume=. 2024 , doi=

    2024

  22. [22]

    npj Computational Materials , volume=

    TransPolymer: a Transformer-based language model for polymer property predictions , author=. npj Computational Materials , volume=. 2023 , publisher=

    2023

  23. [23]

    Nature communications , volume=

    polyBERT: a chemical language model to enable fully machine-driven ultrafast polymer informatics , author=. Nature communications , volume=. 2023 , publisher=

    2023

  24. [24]

    Computational Materials Science , volume=

    Polymer design using genetic algorithm and machine learning , author=. Computational Materials Science , volume=. 2021 , publisher=

    2021

  25. [25]

    Journal of Materials Chemistry C , volume=

    Design of polyimides with targeted glass transition temperature using a graph neural network , author=. Journal of Materials Chemistry C , volume=. 2023 , publisher=

    2023

  26. [26]

    Nature Machine Intelligence , volume=

    Machine learning-aided generative molecular design , author=. Nature Machine Intelligence , volume=. 2024 , publisher=

    2024

  27. [27]

    Nature communications , volume=

    Inverse design of 3d molecular structures with conditional generative neural networks , author=. Nature communications , volume=. 2022 , publisher=

    2022

  28. [28]

    Machine Learning: Science and Technology , volume=

    Self-referencing embedded strings (SELFIES): A 100\ author=. Machine Learning: Science and Technology , volume=. 2020 , publisher=

    2020

  29. [29]

    Npj Computational Materials , volume=

    On-demand reverse design of polymers with PolyTAO , author=. Npj Computational Materials , volume=. 2024 , publisher=

    2024

  30. [30]

    Soft Matter , volume=

    Data-driven algorithms for inverse design of polymers , author=. Soft Matter , volume=. 2021 , publisher=

    2021

  31. [31]

    Journal of Chemical Information and Modeling , volume=

    High-temperature polymer dielectrics designed using an invertible molecular graph generative model , author=. Journal of Chemical Information and Modeling , volume=. 2023 , publisher=

    2023

  32. [32]

    Express Polym

    Molecular dynamics simulation on glass transition temperature of isomeric polyimide , author=. Express Polym. Lett , volume=

  33. [33]

    ACS Polymers Au , volume=

    Emerging trends in machine learning: a polymer perspective , author=. ACS Polymers Au , volume=. 2023 , publisher=

    2023

  34. [34]

    Machine Learning: Science and Technology , volume=

    Chemformer: a pre-trained transformer for computational chemistry , author=. Machine Learning: Science and Technology , volume=. 2022 , publisher=

    2022

  35. [35]

    Nature Machine Intelligence , volume=

    Large-scale chemical language representations capture molecular structure and properties , author=. Nature Machine Intelligence , volume=. 2022 , publisher=

    2022

  36. [36]

    Wireless Communications and Mobile Computing , volume=

    Mol-BERT: an effective molecular representation with BERT for molecular property prediction , author=. Wireless Communications and Mobile Computing , volume=. 2021 , publisher=

    2021

  37. [37]

    Journal of cheminformatics , volume=

    Randomized SMILES strings improve the quality of molecular generative models , author=. Journal of cheminformatics , volume=. 2019 , publisher=

    2019

  38. [38]

    Bioinformatics , volume=

    SimSon: simple contrastive learning of SMILES for molecular property prediction , author=. Bioinformatics , volume=. 2025 , publisher=

    2025

  39. [39]

    Computational Intelligence and Neuroscience , volume=

    A Novel Molecular Representation Learning for Molecular Property Prediction with a Multiple SMILES-Based Augmentation , author=. Computational Intelligence and Neuroscience , volume=. 2022 , publisher=

    2022

  40. [40]

    Briefings in Bioinformatics , volume=

    Complementary multi-modality molecular self-supervised learning via non-overlapping masking for property prediction , author=. Briefings in Bioinformatics , volume=. 2024 , publisher=

    2024

  41. [41]

    npj Computational Materials , volume=

    Property-guided generation of complex polymer topologies using variational autoencoders , author=. npj Computational Materials , volume=. 2024 , publisher=

    2024

  42. [42]

    Journal of Chemical Information and Modeling , volume=

    SMICLR: contrastive learning on multiple molecular representations for semisupervised and unsupervised representation learning , author=. Journal of Chemical Information and Modeling , volume=. 2022 , publisher=

    2022

  43. [43]

    Nature Machine Intelligence , volume=

    Regression transformer enables concurrent sequence regression and generation for molecular language modelling , author=. Nature Machine Intelligence , volume=. 2023 , publisher=

    2023

  44. [44]

    Chemical science , volume=

    Beyond generative models: superfast traversal, optimization, novelty, exploration and discovery (STONED) algorithm for molecules using SELFIES , author=. Chemical science , volume=. 2021 , publisher=

    2021

  45. [45]

    arXiv preprint arXiv:1904.09751 , year=

    The curious case of neural text degeneration , author=. arXiv preprint arXiv:1904.09751 , year=

    Pith/arXiv arXiv 1904

  46. [46]

    Neural computation , volume=

    A learning algorithm for continually running fully recurrent neural networks , author=. Neural computation , volume=. 1989 , publisher=

    1989

  47. [47]

    Peerj computer science , volume=

    The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation , author=. Peerj computer science , volume=. 2021 , publisher=

    2021

  48. [48]

    Journal of chemical information and modeling , volume=

    Extended-connectivity fingerprints , author=. Journal of chemical information and modeling , volume=. 2010 , publisher=

    2010

  49. [49]

    Journal of cheminformatics , volume=

    Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations? , author=. Journal of cheminformatics , volume=. 2015 , publisher=

    2015

  50. [50]

    Journal of chemical information and modeling , volume=

    GuacaMol: benchmarking models for de novo molecular design , author=. Journal of chemical information and modeling , volume=. 2019 , publisher=

    2019

  51. [51]

    npj Computational Materials , volume=

    De novo design of polymer electrolytes using GPT-based and diffusion-based generative models , author=. npj Computational Materials , volume=. 2024 , publisher=

    2024