When Polymer Architecture Changes, Representation Stops Being Preprocessing

TL;DR for operators

A materials team choosing a surrogate model for polymer screening should not treat molecular representation as a preprocessing detail. PolyBench26¹ compares language-model, graph-based, and descriptor-based approaches across 249,962 polymer-property datapoints, eight properties, four polymer architectures, and four evaluation tasks. Graph-based Polymer Chemprop (PC) and Polymer Periodic Graph (PPG) generally achieve the lowest errors on the evaluated in-distribution tasks.

The stronger operational signal appears under structural stress. As alternating-copolymer repeat units become more complex, graph-model errors remain relatively low while descriptor-model errors rise. When random and block copolymer architectures are excluded from training, PC produces roughly one-half to one-third the error of evaluated alternatives that do not natively represent those architectures.

That advantage has a cost. In reported CPU-only timing tests, the graph models take substantially longer to train than polyBERT or Random Forest. The decision for an R&D team is therefore conditional: pay for architecture-aware structural representation when transfer across copolymer forms matters; favor faster descriptor models when retraining speed and compute efficiency dominate.

The boundary matters. The held-out architectures do not contain entirely unseen constituent chemistry, and the benchmark does not cover network, multiblock, gradient, graft, or other more complex polymer classes.

A model can look accurate until the polymer architecture changes

A screening model is valuable because it can rank or filter candidate polymers before a team spends more on simulation or experiment. The difficult question is not whether the model predicts familiar examples well. It is whether the same model remains reliable when the candidate changes in a structurally meaningful way.

PolyBench26 makes that question unusually concrete. It combines experimental, density-functional-theory, and molecular-dynamics-derived targets across homopolymers and alternating, random, and block copolymers. The benchmark then separates four problems that are often collapsed into one headline score: ordinary in-distribution prediction, performance as training data changes, robustness to more complex repeat units, and transfer to polymer architectures withheld from training.

That separation changes how the model ranking should be read. Average error on familiar polymers is only the first test.

Graph models win most of the ordinary accuracy comparisons

On the benchmark’s main in-distribution comparisons, PC and PPG generally deliver the lowest or statistically indistinguishable normalized RMSE across the specialized models.

The scale is visible in the homopolymer examples. For electron affinity, PC and PPG both achieve RMSE around 0.29 eV, compared with 0.39 for polyBERT and 0.43 for Random Forest. For experimental glass-transition temperature, PC records 34.7 °C RMSE, versus 36.7 °C for PPG, 37.9 °C for polyBERT, and 39.7 °C for Random Forest.

The zero-shot GPT comparison is much weaker: GPT-4o records 1.20 eV RMSE for electron affinity and 73.5 °C for glass-transition temperature, while GPT-o3-mini records 0.95 eV and 70.4 °C. Those numbers establish a result about the tested prompting setup, not about language models as a model class. The study does not evaluate polymer-specific fine-tuning.

The more informative question is whether graph models retain their advantage when the conditions become less comfortable.

The graph advantage does not require one universal data regime

The training-size experiments hold validation and test sets fixed while reducing the amount of training data. For experimental and MD-derived properties, graph models remain ahead even in subsets as small as roughly 300 polymers.

But the pattern is not universal. For DFT-derived refractive index, the graph advantage over descriptor-based models becomes apparent only after approximately 1,000 training datapoints.

That exception is useful. It prevents a simplistic conclusion that richer structural representations are automatically more data-efficient. Their value depends partly on the target property and source of the training signal.

For an internal model-selection process, the implication is straightforward: benchmark representation choices at the data scale the team actually possesses. A ranking obtained on the full dataset need not describe the low-data operating regime.

Structural complexity makes representation harder to dismiss

Polymer models first need a machine-readable account of repeating chemical structure. PolyBench26 uses PSMILES for this role across several approaches. Polymer Chemprop instead uses weighted PSMILES, or wPSMILES, which allows composition and connectivity information to be carried into a weighted graph representation.

The distinction becomes more consequential as repeat units become structurally richer. In the benchmark’s alternating-copolymer complexity test, the number of chemically distinct monomers increases from five to ten. PC and PPG maintain low refractive-index prediction error, while errors from polyBERT and Random Forest increase.

This experiment is best read as a representation stress test rather than a second independent thesis. It supports the paper’s proposed mechanism: message-passing graphs preserve local and relational chemical information that global descriptor representations can lose as repeat-unit structure becomes more complicated.

A separate sensitivity result adds an important qualification. Descriptor models are generally less sensitive than the graph models to alternative but structurally equivalent choices of repeat-unit boundaries in PSMILES. Richer representations can preserve more structure while also becoming more dependent on how that structure is encoded.

Architecture transfer is the result that changes the model-selection decision

The strongest practical result comes from withholding random and block copolymer architectures from training while retaining their constituent repeat-unit chemistries in homopolymer and alternating-copolymer training data.

Under this architecture-held-out test, PC has the lowest error. The competing approximations have RMSEs roughly two times higher for polyBERT and roughly three times higher for PPG and Random Forest.

The comparison is not symmetric, however. PC is the only evaluated implementation that natively represents all four architectures. For random and block copolymers, the other approaches are evaluated using molar-ratio-weighted averages of repeat-unit predictions.

That is not a reason to discard the result. It identifies the operational issue more precisely: native representation of polymer architecture is itself part of model capability. A system that cannot express the relevant composition and connectivity cannot be expected to recover them solely through a stronger downstream predictor.

For a team screening across copolymer forms, model architecture and chemical representation therefore have to be selected together.

Better structural representation carries a measurable compute bill

The accuracy gains are not free.

On a 1,800-datapoint homopolymer density dataset, the reported CPU-only training times are 274 seconds for PC and 362 seconds for PPG, compared with 37 seconds for polyBERT and 48 seconds for Random Forest.

At 37,133 alternating-copolymer density datapoints, the difference widens: PC takes 34,218 seconds and PPG 35,221 seconds, while polyBERT takes 497 seconds and Random Forest 4,554 seconds.

These timings come from one Intel Sapphire Rapids CPU core without GPU acceleration, so they should not be treated as universal infrastructure benchmarks. They do establish the direction and magnitude of the tradeoff under a controlled comparison.

Cognaptus inference: an R&D workflow that retrains frequently, explores many target definitions, or operates under tight CPU constraints may rationally accept some predictive loss for faster descriptor-based models. Where one model will support a costly downstream screening campaign across multiple copolymer architectures, the incremental compute required by a graph model may be easier to justify.

What PolyBench26 does not establish

The architecture-transfer result is narrower than “out-of-distribution polymer generalization.” Random and block copolymers are withheld as architectures, but their constituent repeat units remain represented in training. The experiment therefore tests whether a model can reorganize familiar chemical building blocks into an unseen architecture, not whether it can predict entirely novel chemistry.

The benchmark also excludes more structurally complex classes including network, multiblock, gradient, and graft polymers. Its main in-distribution evaluations use random splits rather than chemistry-disjoint or scaffold-like splits. A planned independent hidden test set and online leaderboard are not part of the currently reported evidence.

Some targets also require care. Radius of gyration depends on the finite chain lengths used in the underlying MD simulations and is included mainly as a learning target. The benchmark’s quantum-corrected heat-capacity workflow improves agreement with experimental values substantially, but its calculated $C_v$ is still compared approximately with experimental $C_p$.

These limits define where the benchmark can guide an engineering decision and where additional validation is still required.

Model selection should follow the structural change you expect

PolyBench26 does more than rank polymer-property predictors. It shows why a ranking can change when the stress placed on the representation changes.

For familiar in-distribution prediction, graph models are generally strongest. Across most tested data scales, that advantage persists. When repeat units become more complex, preserving relational structure becomes more valuable. When polymer architecture itself changes, an explicit representation of composition and connectivity becomes decisive in the reported comparison.

The resulting operating rule is bounded but useful: choose the representation for the structural variation the model must survive, not only for its average validation score. Then price the additional accuracy and transfer capability against training cost.

That is a narrower conclusion than declaring graph models universally superior. It is also more actionable.

Cognaptus: Automate the Present, Incubate the Future.