Source-linked AI summary
HiPoly: a hierarchical polymer-native AI framework for property prediction and generative design
Ge Sun, Gervasio Zaldivar, Yuan Tian, Gustavo Perez Lemus, Juhae Park, Dasha Safarian, Ming Han, Juan J. de Pablo
TL;DR
Polymer AI must represent composition, connectivity, and molecular-weight variation across hierarchical structures, but existing approaches do not encode the complete polymer specification in one architecture. HiPoly uses G2RINS and three-level hierarchical graphs to unify prediction, generative design, and MD validation, achieving the highest accuracy for Tg and density while identifying validated PFAS-free candidates. The study demonstrates a shared polymer representation for linking formulation, modeling, generation, and physics-based validation.
Problem
Existing polymer representations do not natively encode complete multi-component specifications, including composition, stochastic connectivity, and molecular weight distribution, within one architecture.
Method
HiPoly uses a three-level hierarchical architecture built on G2RINS to support property prediction, variational generative design, and MD validation through one polymer representation.
Results
HiPoly achieves the highest prediction accuracy for Tg and density and identifies PFAS-free polymer candidates validated through molecular dynamics.
Takeaways & Limitations
The framework links polymer representation, prediction, generation, and physics-based validation across complex polymer chemistries.
Takeaways & Limitations
The molecular-weight dependence of Tg is treated as a learned scalar rather than imposed as a fixed 1/Mn relationship.
Abstract
from arXiv · showhide
Polymeric materials are central to modern technologies, with applications ranging from energy to health and transportation. Although AI has made significant advances in materials discovery, the hierarchical structure of polymers across multiple length scales makes them inherently difficult to represent in a unified and physically meaningful way. Here we introduce HiPoly, a polymer-native AI framework that processes complete polymer descriptions through a three-level hierarchical graph architecture built on the G2RINS representation. HiPoly encodes stochastic inter-monomer connectivity, composition, and molecular weight directly within its architecture, using physically motivated design principles that mirror the multi-scale nature of polymeric systems. The framework establishes an end-to-end AI-driven workflow from experimental formulation data to property prediction, generative molecular design, and physics-based validation through molecular simulations, all unified by a single polymer representation. We demonstrate state-of-the-art prediction accuracy for thermophysical properties of multi-component polymer systems, with ablation studies confirming that each hierarchical design choice contributes independently to model performance. As an example, the generative design pathway is applied here to the discovery of sustainable alternatives to persistent fluorinated polymers, where it is possible to identify and independently validate PFAS-free candidates with target surface-energy properties. This work demonstrates how polymer-native AI can accelerate discovery by linking representation, prediction, and design across complex polymer chemistries.
Introduction
Polymer properties emerge from hierarchical, stochastic structures that monomer-level representations do not fully encode. HiPoly addresses this gap with a G2RINS-based hierarchical architecture supporting prediction, generation, and simulation validation.
- Multi-component polymer properties depend jointly on composition, stochastic inter-monomer connectivity, and molecular weight distribution, creating an experimentally and computationally challenging design space.
- Existing fingerprint, language-model, graph, topology-aware, and physics-guided approaches generally omit part of the complete polymer specification or atomistic chemical detail.
- G2RINS encodes monomer chemistry, mole fractions, stochastic inter-monomer connectivity, and molecular weight distribution in one string that can generate atomistic chain ensembles for MD simulation.
- HiPoly processes polymers across atom, functional-group-motif, and monomer-connectivity resolutions using physically motivated hierarchical design principles.
- The shared architecture supports multi-target property prediction and variational generative design, with candidates validated through MD simulations.
- HiPoly achieves state-of-the-art accuracy for glass transition temperature and density in multi-component systems and identifies MD-validated PFAS-free alternatives with target surface-energy properties.
Results
HiPoly connects formulation data, hierarchical graph encoding, property prediction, generative design, and physics-based validation through G2RINS. It achieves strong prediction performance and generates chemically diverse PFAS-free candidates whose target surface-energy behavior is independently validated by MD.
- HiPoly converts formulation records into G2RINS strings and hierarchical graphs that preserve monomer chemistry, composition, stochastic bonding, and molecular weight distribution.
- The hierarchical encoder feeds either multi-target regression or a shared variational latent space, while G2RINS-generated ensembles provide training labels and independent candidate validation through MD.
- Three physically motivated design principles distinguish HiPoly from flat-graph approaches, including G2RINS-derived stochastic edge weighting for polymerization uncertainty.
- R2 = 0.803 ± 0.147 for Tg and R2 = 0.945 ± 0.024 for density were the highest reported accuracies across the evaluated properties.
- Composition-weighted averaging nearly doubles Tg R2 from 0.36 to 0.70, showing that composition encoding is essential for multi-component systems.
- The model samples 25,000 candidates that fill and expand beyond the reference chemical space, including polymers with PTFE-like surface-energy signatures without fluorinated chemistries.
- Newly validated candidates follow the same predicted-versus-MD γwater relationship as training and test data, supporting generalization to new chemistries.
Discussion
HiPoly combines a complete polymer-native representation with hierarchical physical encoding, prediction, and generative design in one G2RINS-based workflow. Its data-efficient framework achieves strong thermophysical prediction and identifies PFAS-free candidates validated through molecular simulation, while broader datasets and laboratory synthesis remain future steps.
- Representation: HiPoly operates on complete polymer descriptions rather than monomer-level proxies, encoding polymer structure across three hierarchical resolutions.The architecture spans atoms, functional-group motifs, and monomer-level structure.
- Representation: Composition, stochastic bonding, and molecular weight are captured natively within a single end-to-end G2RINS architecture.The representation incorporates weighted composition and probabilistic connectivity alongside hierarchical context.
- Workflow: The framework couples hierarchical encoding with multi-target property prediction and variational generation, connecting formulation data to validated material design.This creates a complete workflow from polymer formulation data through prediction and generative discovery to validation.
- Implications: Data efficiency is a practical advantage for multi-component polymer systems, where labeled data are scarce and expensive to obtain.The authors attribute this efficiency to encoding domain knowledge directly into the architecture rather than relying on large-scale pretraining.
- Implications: The design principles can extend to other graph neural networks and scalar polymer properties, including mechanical, rheological, transport, and dielectric properties.The authors also describe future expansion toward broader polymer families, additional properties, and experimental synthesis and characterization.
- Implications: The PFAS case study demonstrates application to real-world formulation challenges, with polymer-native AI positioned for sustainable packaging and advanced energy-storage materials.The authors frame polymer-native representation, hierarchical physical encoding, and self-consistent validation as principles for future polymer AI models.
Methods
HiPoly formulates polymer property prediction as multi-target regression over complete polymer descriptions, including composition, stochastic connectivity, and molecular-weight distributions. Its G2RINS-derived representation generates realistic chain ensembles that also support molecular-dynamics validation.
- HiPoly predicts continuous properties such as glass transition temperature, density, and interfacial tension from complete polymer descriptions.
- The formulation supports linear and branched polymers with multiple monomer types, variable compositions, and multiple attachment topologies.
- Training accommodates incomplete property labels through NaN-aware loss computation.
- G2RINS encodes polymers as pipe-delimited strings containing monomer SMILES, mole fractions, and directional bond probabilities.
- Bond probabilities and mole fractions are derived from realistic chain ensembles rather than idealized assumptions such as p = 0.5 for random polymers.
- The same chain ensembles provide inputs for molecular-dynamics simulations, aligning AI inputs with atomistic validation.
Hierarchical graph construction
HiPoly constructs a hierarchical polymer graph by decomposing monomers into atoms and chemical motifs, then assembling them into a polymer-level graph. Three message-passing levels preserve molecular detail while applying stochastic weighting specifically to inter-monomer connectivity.
- Functional-group decomposition identifies clusters, pools them into chemically meaningful motifs, constructs motif graphs, and orders edges by depth-first traversal.
- Each monomer fragment yields atom-level and motif-level graphs linked by a cluster-to-atom mapping matrix.
- Fragment motif graphs are merged into a polymer graph carrying bond features, directional weights, degree of polymerization, and composition.
- The encoder passes information bottom-up from atom graphs to motif graphs and then monomer graphs.
- At the monomer level, edge weights equal G2RINS-derived bond probabilities, whereas motif- and atom-level message passing uses unit weights.
- Selective stochastic weighting represents structural uncertainty primarily at inter-monomer connections created by polymerization.
Fragment-level embedding
The fragment-level embedding compresses hierarchical node representations into a fixed-size vector using attention and multiple complementary pooling strategies. This multi-resolution design accommodates properties that depend on different molecular-scale statistics.
- Attention-based pooling converts variable-size hierarchical representations into one fixed-size embedding per monomer fragment.
- A node-level MLP and a pairwise MLP with multi-head self-attention capture individual-node information and inter-node interactions.
- The pooling module concatenates four complementary strategies before projecting to the embedding dimension.
- Attention weights use a query derived from the mean embedding and keys computed for individual nodes.
- The multi-resolution strategy reflects that different material properties may depend on different statistical aspects of molecular representations.
Composition-aware aggregation
HiPoly aggregates fragment embeddings into polymer-level representations using mole-fraction weighting and separately integrates root embeddings. This combines composition information, local chemical detail, and global hierarchical context.
- Fragment embeddings are combined into a polymer embedding through composition-aware aggregation.
- Mole-fraction-weighted summation encodes an inductive bias motivated by approximate linear mixing rules such as the Fox equation.
- Composition is incorporated directly into the architecture rather than treated as an auxiliary metadata feature.
- Root embeddings are composition-weighted, linearly projected, and concatenated with the polymer embedding.
- Ablation experiments show that fragment and root components both contribute to prediction accuracy, with their combination outperforming either alone.
Molecular weight transformation
HiPoly transforms molecular weight and chain length into polymer-level features using a chain-end correction and scaling-aware inputs. The formulation captures finite-chain effects while allowing property-specific behavior across polymer chemistries.
- The polymer-level embedding receives an additive chain-end correction when the degree of polymerization N is specified in the G2RINS string.
- Each chain of N repeat units carries two ends, so chain-end concentration scales as 1/N and shifts bulk properties proportionally to that concentration.
- For N ≫ N0, the correction decays as N0/N and vanishes in the infinite-chain limit.
- A linear readout gives the Fox–Flory form y ≃ y∞ − K/N, with y∞ and K determined by repeat-unit chemistry.
- For power-law properties such as zero-shear viscosity and radius of gyration, the prediction head receives log N, leaving regime crossovers to nonlinear layers.
- N is the number-average degree of polymerization, defined as N = Mn/M0, while full molecular-weight distributions are encoded to capture dispersity and higher-order moments.
Prediction head
HiPoly maps polymer representations to property predictions through a residual MLP and also learns a structured latent space for molecular reconstruction and inverse design. Latent sampling and conditional generation support the PFAS case study.
- Prediction head: Molecular-weight information is mapped to property predictions through a residual MLP whose output dimension equals the number of target properties P.
- Variational pathway: Root vectors are mapped to a latent distribution through learned mean and variance networks in a VAE pathway.
- Variational pathway: The absolute-value constraint ensures σ2 ≤ 1, and latent vectors are sampled using the reparameterization trick.
- Variational pathway: An autoregressive graph decoder reconstructs molecules node-by-node through the motif vocabulary, while KL divergence regularizes the encoder toward smooth, structured latent representations.
- Generative design: The learned latent space supports inverse design through latent-space sampling and conditional generation, demonstrated in the PFAS case study.
Loss function
HiPoly combines masked multi-task property prediction with variational regularization and evaluates the model using structured validation and molecular-dynamics procedures. The workflow accommodates incomplete labels and uses simulation-based interfacial measurements.
- Loss function: The total training objective combines property prediction with variational regularization.
- Loss function: The property loss sums per-property mean squared errors over samples with valid labels, enabling multi-task learning with incomplete polymer-property datasets.
- Loss function: The KL weight is linearly annealed between bounds and additionally scaled by λ, with default λ = 0.01, to balance prediction and regularization.
- Molecular dynamics: Interface simulations join equilibrated organic bulk with water, hexane, and an air-representing vacuum gap, using OPLS and staged thermal equilibration.
- Molecular dynamics: Interfacial tension is calculated with the Test Area 2 method from the free-energy change associated with a virtual constant-volume area perturbation.
- Evaluation: Results use five-fold cross-validation with composite stratification into five quantile bins and a held-out validation set comprising 10% of each training partition.
- Evaluation: Methods in Table 2 use identical splits, while parity plots in Fig. 2c show predictions from a representative fold.
Ablation study design
The ablation study removes one architectural component at a time while keeping the others intact, using Tg as a sensitive representative property. Performance is assessed with standard regression metrics under matched cross-validation splits.
- Ablation design: The study systematically removes exactly one architectural component from the full model while keeping all other components intact.
- Ablation variants: One ablation replaces composition-aware mole-fraction weighting with simple mean pooling over fragment embeddings.
- Ablation variants: Another removes the root-vector pathway that modulates the prediction head, retaining only composition-aggregated fragment embeddings.
- Ablation variants: A third sets G2RINS-derived monomer-graph bond probabilities uniformly to one, removing stochastic edge weighting.
- Evaluation: Tg is used because its sensitivity to cooperative segmental dynamics amplifies the effects of architectural choices.
- Evaluation: Held-out test performance is measured using MAE, RMSE, and R2, with baseline methods evaluated under identical five-fold splits and fold-wise mean and standard deviation.
Data availability
The training polymer datasets were generated using an in-house molecular simulation framework and support the study’s findings. The datasets are available.
- The polymer datasets used for training were generated using an in-house molecular simulation framework.
- These datasets constitute the data supporting the findings of this study.
- The datasets supporting the study’s findings are available.