Periodic Topological Deep Learning for Polymer Design and Discovery
Yasharth Yadav, Tze Kwang Gerald Er, Atsushi Goto, Kelin Xia · 2026
Read Periodic Topological Deep Learning for Polymer Design and Discovery by Yasharth Yadav, Tze Kwang Gerald Er, Atsushi Goto, Kelin Xia (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-26833.txt.
Introduction
Yadav and three colleagues introduce Periodic-TDL, a deep-learning framework for polymer design that captures chain periodicity and many-body interactions missed by single-unit graph models. It builds on periodic Vietoris-Rips complexes and a layered message-passing encoder that moves news from long-range links down to bonds. It beats top rivals on electronic, optical, physical, and thermal tasks. Tests back its predictions that ester-to-amide swaps and backbone methylation raise glass transition points.
Catalog record
| CREATOR | Yasharth Yadav, Tze Kwang Gerald Er, Atsushi Goto, Kelin Xia |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 102157 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.26833
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Periodic Topological Deep Learning for Polymer Design and Discovery Yasharth Yadav, Tze Kwang Gerald Er, Atsushi Goto, Kelin Xia Abstract Polymers underpin applications across energy, healthcare, and materials science, yet their vast chemical space makes systematic discovery challenging. Most machine learning approaches represent polymers as molecular graphs of a single repeating unit, thereby missing both the periodicity of polymer chains and many-body interactions beyond pairwise bonds. We introduce Periodic-TDL, a deep learning framework built on periodic Vietoris-Rips complexes that capture many-body interactions across multiple spatial scales, followed by a hierarchical simplicial message-passing (HSMP) encoder that propagates information from long-range interactions to covalent bonds, yielding representations enriched by higher-order topological features. Periodic-TDL outperforms all state-of-the-art models across polymer property prediction tasks spanning electronic, optical, physical, and thermal targets. Furthermore, we quantitatively validate how ester-to-amide substitution and α \alpha -methylation enhance thermal stability. Using a computationally synthesized dataset of 48,208 structures-generated via systematic substitution of acrylate and acrylamide polymers-we observed a mean T g T_{g} increase of ∼ 55 ∘ \sim 55^{\circ} C for ester-to-amide substitutions and ∼ 14 ∘ \sim 14^{\circ} C for backbone α \alpha -methylation across matched polymer pairs. To verify these predicted trends, we use our Periodic-TDL model to analyze six novel polymer pairs from independent experimental measurements, including three newly synthesized polymers previously unreported in the literature. The experimental data successfully confirmed the model’s predictions. Ultimately, these findings demonstrate that Periodic-TDL captures the underlying physical effects of specific functional group modifications, rather than merely optimizing predictive performance on benchmark datasets. Introduction Polymers underpin modern life across electronics, aerospace, energy, and healthcare 33 ; 56 ; 51 , owing to their tunable thermal, electronic, mechanical, and optical properties. This tunability arises from a virtually infinite chemical space that can be navigated by modifying their molecular architecture and functional group chemistry 61 ; 22 . While the intuition of polymer chemists remains invaluable, the sheer scale of this space renders purely experimental discovery impractical 26 . Polymer informatics has emerged to address this challenge by applying modern tools from data science and machine learning, enabling substantial progress in both property prediction and inverse design, where polymers meeting prescribed requirements are identified 6 ; 20 ; 97 ; 9 ; 47 . Central to any machine learning approach is the choice of how polymer structure is encoded. Early methods derived hand-crafted numerical vectors from local atomic environments, functional groups, and packing descriptors 79 ; 45 ; 23 ; 7 . With the rise of deep learning, the field shifted toward learning directly from machine-readable molecular representations, where numerical vectors emerge as an implicit outcome of the neural encoder. This shift places the burden of accuracy squarely on the representation itself, which must provide a faithful and complete description of the molecular structure for the encoder to extract meaningful information. Sequence-based Transformers 98 ; 49 ; 2 treat the SMILES string 96 as a chemical language, pretraining on large unlabelled corpora before fine-tuning on labelled property datasets. Message-passing neural networks 19 ; 88 ; 91 ; 62 ; 5 ; 3 ; 41 ; 75 ; 37 learn from molecular graphs describing covalent connectivity. Multimodal frameworks 74 ; 89 ; 70 ; 43 combine different representation types to compensate for the blind spots of any single modality. Despite this diversity, most deep learning frameworks construct their structural input from a single, finite repeating unit. As a result, they effectively treat polymers as small molecules rather than as macromolecules, creating a fundamental representational bottleneck that remains largely unaddressed 31 ; 32 ; 54 . Unlike small molecules with a fixed number of atoms, polymers consist of periodic chains of repeating chemical subunits that can extend to millions of atoms, with potentially complex architectures 99 . A faithful representation must therefore structural relationships that span adjacent repeating units in a polymer. Several works have recognised this and proposed periodic graph representations as a more suitable alternative. The wDMPNN framework 3 , the periodic polymer graph 5 , and PolyGNN 37 all construct graph representations that extend beyond a single repeating unit to capture inter-unit connectivity under periodic boundary conditions. However, these approaches share two fundamental limitations. First, graph representations are inherently restricted to pairwise interactions and cannot capture the many-body structure arising from three or more atoms simultaneously. Second, message passing operates only at the scale of covalent bonds, missing long-range non-covalent interactions that are critical for bulk polymer properties. These limitations call for a richer representational framework that can explicitly encode many-body interactions across multiple spatial scales. More recently, topological deep learning (TDL) has emerged as a principled extension of graph neural networks that addresses their fundamental restriction to pairwise interactions 39 ; 38 ; 69 ; 72 . Rather than representing molecular structure as a graph of atoms and bonds, TDL operates on richer topological domains that can natively encode many-body interactions among three or more elements simultaneously. The most widely studied of these are simplicial complexes, which are generalizations of graphs capable of encoding interactions between multiple atoms through higher-dimensional simplices 17 ; 24 ; 68 . Other structures, including cell complexes 29 , sheaves 40 , and hypergraphs 11 , are also gaining traction. Message passing schemes defined across these structures enable information exchange between objects of varying dimension. This allows representations to incorporate both local and global structural information 14 ; 13 ; 83 ; 34 . These capabilities have demonstrated considerable success in molecular property prediction 16 ; 94 ; 35 ; 4 ; 85 . However, no existing TDL framework incorporates the periodic nature of polymers, where chains are still treated as finite molecular fragments, discarding the key structural aspect that distinguishes polymers from small molecules. Here we introduce Periodic-TDL, a framework that captures both periodicity and higher-order interactions across multiple spatial scales. At its core is the periodic Vietoris–Rips complex representation of polymers, which encodes interactions between atoms across adjacent repeating units under periodic boundary conditions, representing both covalent and non-covalent interactions through higher-dimensional simplices. To learn from this representation, we develop the Hierarchical Simplicial Message Passing (HSMP) encoder, which operates over a nested sequence of simplicial complexes and propagates information from coarser to finer scales in a chemically motivated direction. Notably, at the finest filtration level, HSMP reduces to a standard molecular graph, enabling direct integration into established self-supervised pretraining pipelines 80 ; 91 ; 104 ; 90 ; 105 ; 46 designed for conventional graph encoders. This compatibility is particularly valuable given the scarcity of labelled polymer data. To equip higher-order simplices with meaningful features, we further introduce curvature-based featurization using Forman’s discretization of Ricci curvature 29 , which generalizes naturally to simplices of any dimension. Next, we pretrained Periodic-TDL on one million unlabelled polymers and evaluated its performance across nine polymer properties spanning electronic, optical, physical, and thermal targets. Finally, we applied Periodic-TDL to a computationally generated dataset of 48208 systematically substituted polymers from acrylate and acrylamide families. Specifically, we studied the effects of ester-to-amide substitution and backbone \alpha -methylation on the glass transition temperature ( T_{g} ) of these polymers. To ensure the chemical credibility of our model, we compared the predicted T_{g} changes against six polymer pairs spanning eight polymers