1. The short version / before you read / where this sits in the series

• A polymer is not one structure per name: chain length has a distribution, and arrangement, branching, end groups and how it was made all matter. The BigSMILES paper calls polymers "intrinsically stochastic molecules" Sourced [Ref. 2].
• The data are lopsided: of 14,423 homopolymers in PoLyInfo (as of 2019), about 41% have glass-transition data, about 22% melting points, and thermal conductivity only 28 (about 0.19%) Sourced / Our calculation [Ref. 5].
• Inverse design has gone as far as synthesis and measurement: in the design of polyamides of high thermal conductivity, three were synthesized from thousands of candidates and measured at 0.18 to 0.41 W/mK. But the prediction was a guide built on 28 training data Sourced [Ref. 5].

2. How polymer data differ from small-molecule data

(1) Difficulty 1: one sample is a collection of chains of different lengths

A sample of ethanol contains only molecules of the same structure. So the name "ethanol" fixes one structure and one molecular weight. A polymer sample, however, is a collection of chains of different lengths (a distribution). Even for one sample, the number-average molecular weight (Mn) and the weight-average molecular weight (Mw) differ, and their ratio (dispersity, Đ) expresses the breadth of the distribution (details are in the explainer on polymer basics; IUPAC recommended the term "dispersity" for Mw/Mn [Ref. 16]).

Fig. 2 shows an example of the most probable distribution of step-growth polymerization (the equation Flory derived in 1936 [Ref. 22]). At a conversion of 0.98, the average chain length (number-average) is 50 but the weight-average is 99, so chains an order of magnitude apart in length are mixed together. Our calculation. What matters here is how data are recorded. A bare "molecular weight 50,000" means different things depending on whether it is Mn or Mw and how it was measured (for SEC, against what it was calibrated).

(2) Difficulty 2: the same composition can be arranged differently

Consider a copolymer made from two monomers, A and B. Even at a composition of A to B of 1:1, the arrangement can be alternating, random, block and so on (Fig. 3).

For a chain of length 20 there are about 1.05 million arrangements, and for length 50 about 1.1×10¹⁵ (Our calculation; idealized numbers). The arrangements actually obtained in synthesis are limited by the reactivity of the monomers and so on. Even so, the composition "A to B of 1:1" alone does not fix the structure. In addition, branching, end groups and tacticity (the order of the orientation of substituents) affect properties.

(3) Difficulty 3: properties also depend on "how it was made"

Even for the same structure, the properties of a polymer change with molecular weight, degree of crystallinity, molding conditions, thermal history, moisture and additives. The paper by Wu et al. says that the thermal-conductivity data in PoLyInfo varied widely even for the same polymer, and gives differences in processing, in higher-order structure and in measurement conditions as possible causes Sourced [Ref. 5]. The same paper also points out that thermal conductivity changes greatly when molecules orient during film molding or fiber spinning, and that most such variations are not recorded in the database [Ref. 5].

3. How to write structure — from repeat unit to population of chains to sample

To give a polymer to AI, its structure has to become numbers or a character string. For polymers, what is written splits into four levels (Fig. 4).

• Repeat-unit representation (level 2): a form that extends SMILES for small molecules (a way of writing structure as a string) to the repeat unit is used by many platforms. Polymer Genome is a web platform that takes the SMILES of a polymer's repeat unit and predicts properties. Its official site states that "it is not a database but an informatics platform that predicts properties with pre-built machine-learning models" Sourced [Ref. 3, 20].
• Representing a population of chains (level 3): BigSMILES extends SMILES by introducing a "stochastic object" for the intrinsically stochastic part. The idea is that one string refers to a group of molecules with a distribution of structures. The paper reports a pilot test converting a standard glass-transition dataset (Bicerano, 2002) into BigSMILES Sourced [Ref. 2]. Aldeghi and Coley proposed a graph representation that treats a polymer as an "ensemble of molecules" and a matching graph neural network; they report that it captures chain architecture, monomer stoichiometry and degree of polymerization, and is more accurate than off-the-shelf cheminformatics methods. For this work they built a dataset of calculated electron affinities and ionization potentials for more than 40,000 polymers Sourced [Ref. 4].
• Representing a sample (level 4): CRIPT (Community Resource for Innovation in Polymer Technology) was proposed as a data model to handle the complexity of defining a polymer's structure and the intricacies of characterizing material properties Sourced [Ref. 7].

4. Databases — PoLyInfo and others

(1) PoLyInfo: a collection of experimental data gathered from the literature

PoLyInfo is the polymer database of the National Institute for Materials Science (NIMS). According to the official site, it gathers the data needed to design polymeric materials from the academic literature and organizes them, recording properties, chemical structures, names (including IUPAC-based names), the molding method of the measured sample, measurement conditions, source monomers and polymerization methods in mutual relation. It covers about 100 types of properties, including thermal, electrical and mechanical ones, and publishes homopolymers, copolymers, polymer blends, composites and compounds Sourced [Ref. 1].

Source: the PoLyInfo official site (checked 23 October 2026). Sourced.

The official site also gives cautions on use. Use requires user registration for a DICE account, domain registration of the e-mail address and an application for service use. Mass downloading and web scraping (automatic retrieval) are prohibited by the MatNavi Service Terms of Use, whether manual or mechanical. Provision in dataset form, including unpublished data, is for a fee [Ref. 1]. Check the terms of use before using it to train your own AI (we did not check the text of the terms).

(2) Other resources: databases, generated data and calculated data

Counts and performance are as reported in each paper's abstract and text or on the site. Sourced. They do not rank the resources.

5. Data imbalance and the effect of molecular weight — our calculation

(1) Data volume differs by orders of magnitude between properties

The paper by Wu et al. (2019) gives the data volume by property for 14,423 unique homopolymers in PoLyInfo. There are 5,917 for glass-transition temperature (Tg) and 3,234 for melting temperature (Tm), but for thermal conductivity near room temperature, only 28 homopolymers (322 measured values) Sourced [Ref. 5]. By our calculation, out of 14,423 that is about 41% for Tg, about 22% for Tm and about 0.19% for thermal conductivity (Our calculation).

This imbalance shapes how models are built. For thermal conductivity, where data are scarce, Wu et al. used transfer learning: they pretrained on Tg, Tm and density (PoLyInfo) and on specific heat (QM9), then fine-tuned with the small thermal-conductivity data. They report that the best transferred model had a mean absolute error of 0.0204 W/mK, smaller than the 0.0327 W/mK of a random forest trained directly on the 28 points Sourced [Ref. 5]. At the same time, the paper says validation was done only in the small region spanned by the 28 compounds, so uncertainty remains about generalization over a wider materials space [Ref. 5].

(2) Even with the same repeat unit, properties change with molecular weight

Let us look at the effect of molecular weight with the Flory–Fox equation found in textbooks (Tg = Tg∞ − K/Mn) [Ref. 15]. Fox and Flory reported in 1950 that the glass-transition temperature of polystyrene changes with molecular weight (we checked this at the level of the title). Here we calculate with illustrative assumptions (Tg∞ = 373 K, K = 1.0×10⁵ K·g/mol) (Fig. 6).

Under this assumption, Tg is 353 K at Mn = 5,000, 363 K at 10,000, 371 K at 50,000 and 372 K at 100,000 (Our calculation). In the range of small molecular weight, Tg changes a lot; as it grows, Tg levels off. A model that takes only the repeat unit as input treats samples with Mn of 5,000 and 100,000 as the same thing, so this difference of about 20 K stays as "scatter". This is one reason to record molecular weight with the data (or to use only data in the range of sufficiently large molecular weight).

6. Progress in prediction models — fingerprints, multi-task learning, language models, graphs

Models that predict properties from the repeat unit have advanced a great deal in a few years. Here we line up the main directions from the abstracts (the figures are as reported in each paper; because conditions differ, they cannot be used to rank the methods).

• Multi-task learning (Kuenneth et al., 2021): they gave data on 36 properties of more than 13,000 polymers to a deep-learning multi-task architecture. Because it can use correlations among properties, they report it is accurate, efficient and scalable compared with conventional single-task models. It is also interpretable, and chemical rules for how features control trends in property values emerge Sourced [Ref. 11].
• Graph neural networks (Gurnani et al., 2023): they showed that having the model learn important features from the repeat unit, instead of making them by hand, speeds up feature extraction by 1 to 2 orders of magnitude relative to handcrafted methods without compromising accuracy Sourced [Ref. 13].
• The chemical language model polyBERT (Kuenneth and Ramprasad, 2023): it treats polymer structure as a "chemical language" and is reported to outstrip the best handcrafted-fingerprint methods in speed by two orders of magnitude while preserving accuracy. For training, 80 million of 100 million hypothetical PSMILES, made by breaking down and recombining 13,766 known polymers, were used Sourced [Ref. 12].
• Graphs of molecular ensembles (Aldeghi and Coley, 2022): the method, seen in the previous section, that captures chain architecture, monomer ratio and degree of polymerization [Ref. 4].

7. Inverse design in practice — propose by calculation, then synthesize and measure

Inverse design is the reverse problem of proposing a "structure" from a "desired property". For polymers, whether the proposed structure can be synthesized is especially important. Here we set out two examples that went as far as synthesis and measurement, and examples of calculation and generation.

(1) Polymers of high thermal conductivity (Wu et al., 2019)

Wu et al. built a forward model from PoLyInfo's Tg and Tm data, inverted it, and generated structures of hypothetical polymers in the desired region of Tg and Tm. A thermal-conductivity model was built by transfer learning and used to screen the generated candidates. According to the abstract, they identified thousands of promising hypothetical polymers, selected three of them for synthesis and polymerization of monomers on the grounds of synthetic accessibility and ease of processing, and the synthesized polymers reached thermal conductivities of 0.18 to 0.41 W/mK, comparable to state-of-the-art polymers among non-composite thermoplastics Sourced [Ref. 5].

• The flow in the text: from 1,000 designed candidates, they narrowed to 24 by resemblance to liquid-crystalline-polymer (LCP) structures, synthesizability (SA score) and ease of processing (Tg of 300 or less; the unit was not checked here); routes could be found for three kinds of polyamide (a wholly aromatic polyamide, an aromatic polyhydrazide and an aliphatic–aromatic polyamide), and they were successfully synthesized [Ref. 5].
• The authors state that the predicted thermal conductivities of the three chosen were not particularly high, because they placed importance on synthesizability and ease of processing [Ref. 5].
• The authors state that they built the workflow on the premise that "predicted properties differ from reality" and used the computationally designed candidates as a guide for chemists' decisions [Ref. 5].

(2) Dielectrics for high temperatures (Gurnani et al., 2024)

Gurnani et al. combined AI, established polymer chemistry and molecular engineering to search for dielectrics in the polynorbornene and polyimide families. For electrostatic-capacitor dielectrics, there is demand for both high energy density and high thermal stability. They report that many of the dielectrics they found show high thermal stability and high energy density over a wide temperature range, and that one shows an energy density of 8.3 J/cc at 200 °C, 11 times the value of any commercially available polymer dielectric at that temperature Sourced [Ref. 6]. According to the text, they generated 26,858 structures of polymers that can be made by ROMP (ring-opening metathesis polymerization), screened them by machine-learning-predicted Tg, dielectric constant and band gap, and, from the top twelve candidates, chose five as synthetic targets on the grounds of raw-material cost and an expert evaluation of synthesizability. Of the five, four were prepared as high-molecular-weight samples and evaluated as films, and one could not be polymerized Sourced [Ref. 6]. For the polyimide family, the text presents them as computational proposals whose evaluation by experiment it hopes the research community will carry out [Ref. 6].

(3) Examples of calculation and generation: RadonPy and OMG

RadonPy found, from more than 1,000 amorphous polymers by molecular dynamics, eight with thermal conductivity above 0.4 W/mK [Ref. 10]. But these are calculated values. OMG proposed generating polymers limited to synthesizable chemistry, but it stops at a proof of concept (designing a target solubility) and includes no synthesis or measurement [Ref. 8].

8. Telling prediction from experiment — the ladder of confirmation

We line up the reports so far by "how far they were confirmed". For polymers, it helps to divide the stages into five rungs (this article's own arrangement).

1. Calculated prediction or candidates: properties or candidates produced by models or calculation
2. Polymerized, structure confirmed: actually synthesize and confirm molecular weight and structure
3. Property measured: measure thermal conductivity and so on under set conditions
4. Parts and real-use conditions: molding, long-term durability, evaluation in real parts
5. Independent check, mass production: replication by other groups, reproduction at scale

As Fig. 9 shows, PI1M and OMG are at rung 1 (candidates, prediction), and RadonPy adds comparison with literature experimental values to rung 1. Wu et al. and Gurnani et al. went through rungs 1 to 3 (synthesized and measured). No report among those covered here could be confirmed through rung 4 (parts and real-use conditions) and rung 5 (independent verification, mass production). For polymers, moldability in real parts, long-term degradation and cost are asked separately from the material's properties, so there is a large wall between rung 3 and rung 4 (our commentary).

9. A materials engineer's view

What can be left to AI, and where expert judgment is needed

The table is this article's own arrangement (commentary).

10. Limits, what could not be confirmed, and summary

(1) Papers cited from the abstract or title only

Kuenneth et al. (2021), Gurnani et al. (2023), Hayashi et al. (RadonPy), Walsh et al. (CRIPT), Kim et al. (OMG), Ma and Luo (PI1M) and Chen et al. (review) were checked mainly from their abstracts. Fox and Flory (1950), Flory (1936) and Stepto (the IUPAC recommendation of 2009) were checked at the level of title and bibliographic data. Wu et al. (2019), Kuenneth and Ramprasad (polyBERT) and Gurnani et al. (2024) were checked in their abstracts and parts of the text.

(2) Terms of use of the databases

The official site of PoLyInfo states that registration is required, that mass downloading and scraping are prohibited, and that dataset form is for a fee. We did not check the text of the terms of use. We did not check the licenses (whether commercial use is allowed and so on) of the data and models of PI1M, RadonPy, OMG, CRIPT and Polymer Genome.

(3) Current data volume by property

The per-property counts in Fig. 5 are as stated in the paper of Wu et al. (2019); we did not check the current per-property breakdown of PoLyInfo. What we could confirm on the official site is only the overall counts (as of 30 September 2026).

(4) What we do not cover

We do not cover designing polymers for recycling and degradability, handling data on blends and composites, extracting data from the literature with large language models, or company case studies of polymer design.

(5) Developments in Japan

We did not survey in detail the overall picture of NIMS's data infrastructure (MDR and others) or the efforts of institutions and companies in Japan.

11. How to read the yardsticks and standards

How good MI for polymers is depends on what is taken as the correct answer (an experimental value from the literature or a calculated value), whether measurement methods are aligned (the type of molecular-weight average, DSC conditions), how the evaluation data are split (by family?) and how far it was confirmed (Section 8). There are international standards for measuring polymers, but we could not find an international standard specific to MI for polymers (prediction models or inverse design). Here is what we could confirm (checked 23 October 2026).

The arrangement of the table is Our calculation (arrangement). The numbers, titles, editions and status of ISO standards were checked on the official ISO pages (23 October 2026). The texts of standards are paid, so only their scope is noted.

Cautions when reading (for materials engineers)

• A ranking in a paper is not the ability on your polymer: the accuracy in a paper is a value within the range of its training data. Re-measure on your own family (polyamides, elastomers and so on) and with your own way of measuring.
• Tell prediction from experimental confirmation: always check which rung of the ladder in Section 8 a result is at. "Designed by AI" is often a rung 1 story.
• Align the way of measuring: molecular weight (the type of average and calibration) and Tg (DSC conditions) change with the way of measuring. The standards (ISO 16014-1, ISO 11357-2) set the procedure but are not criteria for AI.
• How strong a principle or standard is: the FAIR principles and BigSMILES are recommendations or proposals, and most ISO standards are voluntary (the latter is our commentary). A database's terms of use are conditions that must be observed.

12. Glossary

13. References

1. National Institute for Materials Science (NIMS) "Polymer Database (PoLyInfo)" (official site; counts as of 30 September 2026; checked 23 October 2026) — polymer.nims.go.jp/en/
2. Lin T.-S. et al. "BigSMILES: A Structurally-Based Line Notation for Describing Macromolecules", ACS Cent. Sci. 5, 1523–1531 (2019), DOI 10.1021/acscentsci.9b00476 (open access) — doi.org/10.1021/acscentsci.9b00476
3. Kim C. et al. "Polymer Genome: A Data-Powered Polymer Informatics Platform for Property Predictions", J. Phys. Chem. C 122, 17575–17585 (2018), DOI 10.1021/acs.jpcc.8b02913 (checked at the level of title and bibliographic data) — doi.org/10.1021/acs.jpcc.8b02913
4. Aldeghi M., Coley C. W. "A graph representation of molecular ensembles for polymer property prediction", Chem. Sci. 13, 10486–10498 (2022), DOI 10.1039/d2sc02839e (open access) — doi.org/10.1039/d2sc02839e
5. Wu S. et al. "Machine-learning-assisted discovery of polymers with high thermal conductivity using a molecular design algorithm", npj Comput. Mater. 5, 66 (2019), DOI 10.1038/s41524-019-0203-2 (open access) — nature.com/articles/s41524-019-0203-2
6. Gurnani R. et al. "AI-assisted discovery of high-temperature dielectrics for energy storage", Nat. Commun. 15, 6107 (2024), DOI 10.1038/s41467-024-50413-x (open access; an author-affiliation correction was published in August 2024) — nature.com/articles/s41467-024-50413-x
7. Walsh D. J. et al. "Community Resource for Innovation in Polymer Technology (CRIPT): A Scalable Polymer Material Data Structure", ACS Cent. Sci. 9, 330–338 (2023), DOI 10.1021/acscentsci.3c00011 — doi.org/10.1021/acscentsci.3c00011
8. Kim S. et al. "Open Macromolecular Genome: Generative Design of Synthetically Accessible Polymers", ACS Polym. Au 3, 318–330 (2023), DOI 10.1021/acspolymersau.3c00003 — doi.org/10.1021/acspolymersau.3c00003
9. Ma R., Luo T. "PI1M: A Benchmark Database for Polymer Informatics", J. Chem. Inf. Model. 60, 4684–4690 (2020), DOI 10.1021/acs.jcim.0c00726 — doi.org/10.1021/acs.jcim.0c00726
10. Hayashi Y. et al. "RadonPy: automated physical property calculation using all-atom classical molecular dynamics simulations for polymer informatics", npj Comput. Mater. 8, 222 (2022), DOI 10.1038/s41524-022-00906-4 (open access) — nature.com/articles/s41524-022-00906-4
11. Kuenneth C. et al. "Polymer informatics with multi-task learning", Patterns 2, 100238 (2021), DOI 10.1016/j.patter.2021.100238 — doi.org/10.1016/j.patter.2021.100238
12. Kuenneth C., Ramprasad R. "polyBERT: a chemical language model to enable fully machine-driven ultrafast polymer informatics", Nat. Commun. 14, 4099 (2023), DOI 10.1038/s41467-023-39868-6 (open access) — nature.com/articles/s41467-023-39868-6
13. Gurnani R. et al. "Polymer Informatics at Scale with Multitask Graph Neural Networks", Chem. Mater. 35, 1560–1567 (2023), DOI 10.1021/acs.chemmater.2c02991 — doi.org/10.1021/acs.chemmater.2c02991
14. Chen L. et al. "Polymer informatics: Current status and critical next steps", Mater. Sci. Eng. R 144, 100595 (2021), DOI 10.1016/j.mser.2020.100595 (abstract checked at arXiv:2011.00508) — arxiv.org/abs/2011.00508
15. Fox T. G., Flory P. J. "Second-Order Transition Temperatures and Related Properties of Polystyrene. I. Influence of Molecular Weight", J. Appl. Phys. 21, 581–591 (1950), DOI 10.1063/1.1699711 (checked at the level of title and bibliographic data) — doi.org/10.1063/1.1699711
16. Stepto R. F. T. "Dispersity in polymer science (IUPAC Recommendations 2009)", Pure Appl. Chem. 81, 351–353 (2009), DOI 10.1351/PAC-REC-08-05-02 — doi.org/10.1351/PAC-REC-08-05-02
17. ISO "ISO 16014-1:2019 Plastics — Determination of average molecular weight and molecular weight distribution of polymers using size-exclusion chromatography — Part 1: General principles" (ISO page) — iso.org/standard/74995.html
18. ISO "ISO 11357-2:2020 Plastics — Differential scanning calorimetry (DSC) — Part 2: Determination of glass transition temperature and step height" (ISO page) — iso.org/standard/77310.html
19. ISO "ISO 10350-1:2025 Plastics — Acquisition and presentation of comparable single-point data — Part 1: Moulding materials" (ISO page) — iso.org/standard/85826.html
20. Ramprasad Research Group "Polymer Genome" (official site; we confirmed its statement that it is not a database but a prediction platform) — polymergenome.org
21. Wilkinson M. D. et al. "The FAIR Guiding Principles for scientific data management and stewardship", Sci. Data 3, 160018 (2016), DOI 10.1038/sdata.2016.18 — doi.org/10.1038/sdata.2016.18
22. Flory P. J. "Molecular size distribution in linear condensation polymers", J. Am. Chem. Soc. 58, 1877–1885 (1936), DOI 10.1021/ja01301a016 (checked at the level of title and bibliographic data) — doi.org/10.1021/ja01301a016

14. Claim-to-source audit

Last updated 23 October 2026. Sources are limited to primary material (peer-reviewed papers, the official sites of PoLyInfo and Polymer Genome, and official ISO pages). The texts of ISO standards are paid and could not be checked. The calculation examples are made by us with theoretical formulas and illustrative assumptions, not measured values of real polymers. An international standard specific to MI for polymers, the current PoLyInfo counts by property, the licenses of each dataset and model, the text of the PoLyInfo terms of use, and reports confirmed through real parts or independent verification could not be confirmed and are not stated here. All figures are explanatory concept graphics or drawings that include our calculations. Figs. 1 and 8 are ImageGen-generated conceptual illustrations, and none of them shows a real instrument or measurement result.