Closed-loop Polymer Structure–Property Predictor

• Closed-loop polymer structure–property prediction is an integrated system that automates polymer discovery by linking generative modeling, machine learning, and feasibility scoring.
• Its methodology leverages diverse molecular representations like SMILES, BigSMILES, and graph-based ensembles to capture polymer characteristics and enhance predictive accuracy.
• The system iteratively refines candidate structures through property prediction and LLM-guided edits, effectively accelerating laboratory validation and material innovation.

A closed-loop polymer structure–property predictor is an integrated computational framework for automated polymer discovery, connecting generative modeling, machine-learning-based property prediction, synthetic feasibility assessment, and iterative refinement in a feedback-driven cycle. Such systems aim to resolve the historic bottleneck of trial-and-error experimentation in polymer development by autonomously proposing, evaluating, and optimizing candidate polymer structures with respect to specified property targets, all with a central focus on laboratory applicability and robust representational fidelity.

1. System Architectures and Principal Components

Closed-loop predictors universally comprise three or more modular engines operating under programmatic or agentic orchestration:

• Reasoning Core (LLM/Logic controller): Interprets user queries, assigns workflow steps, and manages tool invocation. Implemented using terminal-integrated LLMs such as PolyAgent’s Model Context Protocol (MCP)-connected core (Nigam et al., 23 Jan 2026).
• Structure Generator: Employs reaction-aware VAEs, Transformer-based decoders, or evolutionary algorithms to sample new CRU SMILES/P-SMILES. MoleculeChef VAE architecture exemplifies property-conditional generation via latent space traversal.
• Property Predictor: Trained ML surrogate, typically Transformer-based (e.g. TransPolymer), GNN, or multimodal fusion network, returns predicted property vector $\hat{y}$ from SMILES, sequence, or reconstructed graph representation.
• Structure Modifier: Suggests localized edits to candidate polymers (side-chain substitutions, branching, etc.) and validates improvement via re-prediction.
• Feasibility Scoring: Synthetic Accessibility (SA) and Synthetic Complexity (SC) scores rank candidates by ease of synthesis and process complexity (Nigam et al., 23 Jan 2026).

This architecture is implemented in iterative cycles: candidate structure generation → property prediction → feasibility scoring → local refinement → loop until convergence or resource exhaustion.

2. Molecular Representations and Data Encoding

Choice of representation is crucial for predictive performance and generative validity:

• SMILES and P-SMILES: Linearized notation for constitutional repeating unit; P-SMILES adds “*” for polymerization sites, enabling robust Transformer tokenization (Nigam et al., 23 Jan 2026).
• BigSMILES/CurlySMILES: Encoded polymerizable blocks and side architectures, important for capturing copolymer and blend compositions (Li, 30 Oct 2025).
• PSELFIES (polyBART): Converts polymer repeat units to canonical SELFIES via artificial terminal atoms, allowing transfer learning from molecular LLMs (Savit et al., 21 May 2025).
• Graph-Based Ensembles: Directed weighted graph formulation (wD-MPNN) models monomer stoichiometry, bond formation probabilities, and chain architecture (Aldeghi et al., 2022).
• Simplicial Complexes (Mol-TDL): Multiscale Vietoris–Rips filtration enables topological deep learning over atoms, edges, and higher-order features (Shen et al., 2024).
• Infinite Polymer Sequence (MIPS): Star-linking graphs encode infinite repetition and backbone embeddings handle ring substructures (Wang et al., 27 Jul 2025).

Representational choice directly impacts model invariance (e.g., RSIT in MIPS), capacity to handle blends/copolymers, and downstream gradient or acquisition step fidelity.

3. Property Prediction Models

Regression surrogates draw on task-appropriate architectures:

• Transformer-Based Models: RoBERTa/SMILES-BERT encoders with regression heads are fine-tuned on large curated datasets enforcing mean-squared error loss, e.g., PolyAgent’s TransPolymer model with performance $R^2$ scores ranging from 0.69 (conductivity) to 0.93 (bandgap) (Nigam et al., 23 Jan 2026).
• Multimodal Fusion: MMPolymer combines sequence and 3D geometric features, aligning via cross-modal contrastive learning and SE(3)-equivariant layers (Wang et al., 2024).
• GNNs and MPNNs: Weighted directed architectures (wD-MPNN) incorporate ensemble information, copolymer composition, and stochastic chain formation, achieving $R^2$ = 0.997–1.0 (EA/IP) (Aldeghi et al., 2022).
• Topological Deep Learning: Multiscale message-passing over k-simplices in Mol-TDL yields improved predictive accuracy by integrating non-covalent and higher-order interactions (Shen et al., 2024).
• Gaussian Process (polyBART): Encodes learned latent spaces for regression via RBF kernels on PCA-reduced structure embeddings (Savit et al., 21 May 2025).

All models report standard regression accuracy metrics (RMSE, MAE, $R^2$), cross-validation protocols, uncertainty quantification, and task-wise performance. Multi-task and transfer learning strategies are used to extend prediction coverage with minimal data.

4. Generative and Inverse Design Algorithms

Polymer candidate generation employs multiple modalities:

• VAE-Based Gradient Optimization: Property head-augmented VAEs traverse latent $z$ by gradient steps to maximize or minimize $|p(z) - t|$, yielding candidates nearest to property objectives (Nigam et al., 23 Jan 2026, Li, 30 Oct 2025).
• BART/SELFIES Decoding: polyBART enables bidirectional generation; property-aligned latent codes are sampled, decoded, and validity-filtered at scale (Savit et al., 21 May 2025).
• Genetic Algorithms/BO: Selection, crossover, and mutation on population chromosomes; Bayesian Optimization uses GP surrogates, expected improvement acquisition functions, and hybrid multi-objective front identification (Li, 30 Oct 2025, Shukla et al., 2023).
• LLM-Guided Edits: Natural language agents propose context-specific modifications to SMILES strings, validated against predictive surrogates for monotonic property improvement (Nigam et al., 23 Jan 2026).

Iterative improvement cycles continue until convergence criteria are met (max iterations, lack of improvement, early stop).

5. Synthetic Accessibility and Complexity Scoring

Synthetic feasibility is quantified via rigorous scoring metrics embedded in ranking functions:

Score Type	Calculation	Range/Thresholds
Synthetic Accessibility (SA)	$SA(m) = f_{\text{frag}}(m) + p_{\text{size}}(m)$, fragment frequency and complexity penalties; Ertl et al. 2009	1 (easy) – 10 (hard); SA $\leq 3$ = synth.
Synthetic Complexity (SC)	Learned regressor, SMILES $\rightarrow$ [1,5], pairwise ranking; Coley et al. 2018	SC $\approx 1$ = simple, $\geq 4$ = complex
PolymerGenome SA	Empirical, thresholded for online tool compatibility	SA $\leq 6$ preference in polyBART

Candidates are jointly ranked by $-|ŷ(m) - t| - λ_1 SA(m) - λ_2 SC(m)$ or analogous scalarizations (Nigam et al., 23 Jan 2026, Savit et al., 21 May 2025). This mitigates impractical suggestions and aligns design outcomes with experimentally accessible structures.

6. Closed-Loop Design Workflows

Closed-loop workflows are orchestrated via stepwise logic:

1. Specification: User submits target property profile (e.g., log conductivity, synthetic ease).
2. Generation: Generator proposes $n$ candidate SMILES/P-SMILES, filtered by preliminary SA/SC thresholds.
3. Prediction: Surrogate model predicts relevant properties for each candidate.
4. Scoring/Selection: Rank candidates on $\Delta$ from target, SA/SC, multi-objective criteria.
5. Refinement: Top candidates are modified via RL, LLM edits, or gradient optimization; new variants are checked.
6. Iteration: Loop returns to prediction, stopping on satisfaction of all thresholds or exhaustion of resources.

A canonical example is provided in the PolyAgent system: a user requests a polymer with log conductivity $-5.3$ and low complexity; the system cycles through candidate suggestion, prediction, scoring, and local edit refinement, outputting a concise synthesis report upon candidate acceptance (Nigam et al., 23 Jan 2026). Similar workflows underpin polyBART (Savit et al., 21 May 2025), MMPolymer (Wang et al., 2024), MIPS (Wang et al., 27 Jul 2025), and wD-MPNN/GA pipelines (Aldeghi et al., 2022, Li, 30 Oct 2025).

7. Performance, Validation, and Laboratory Applicability

Benchmarking against established computational (Polymer Genome) and experimental targets shows closed-loop predictors reach and often exceed state-of-the-art accuracy:

Property	PolyAgent Prediction	Polymer Genome
Electron Affinity (Eea)	0.99 eV	1.0 ± 0.4 eV
Bulk Bandgap (Egb)	5.68 eV	6.4 ± 0.5 eV
Dielectric Constant (EPS)	3.44	4.54 ± 0.68
OPV Efficiency (PCE)	4.73%	n/a

polyBART has demonstrated experimental synthesis and property validation within $5$–$10$ K of prediction targets for $T_g$ and $T_d$ (Savit et al., 21 May 2025). Similar empirical matches have been reported in MMPolymer (Wang et al., 2024) and MIPS (Wang et al., 27 Jul 2025).

Robustness is tested via cross-validation strategies (k-fold, LOCO-CV), as well as uncertainty estimation (ensembles, MC-dropout) and continuous learning via retraining on new experimental data.

8. Future Directions, Challenges, and Integration

Emerging research emphasizes:

• Multiscale and Multimodal Representation: Integration of sequence, 3D, and topological features improves predictive accuracy and transferability to unseen polymer chemistries (Wang et al., 2024, Shen et al., 2024).
• Physics-Informed ML: Augment property predictors/ecodes with DFT/CGMD features (Li, 30 Oct 2025).
• Standardization and Interpretability: Unified data schemas (BigSMILES, CRIPT), feature attribution (SHAP), and transparent scoring (Li, 30 Oct 2025).
• Automated Laboratory Integration: Terminal/robotics/free-coding interfaces for rapid experimental synthesis and evaluation (Nigam et al., 23 Jan 2026).

Current limitations include domain extrapolation, practical verification in non-standard chemistries, and balancing synthetic feasibility with property optimization. A plausible implication is that continual improvement in representation, agent logic, and benchmarking will be needed to generalize closed-loop predictors across polymer classes and application domains.

These systems now offer accessible, precise, and laboratory-ready frameworks for the rational discovery of functional polymers, significantly accelerating the pathway from target specifications to validated material design.