Autonomous Multi-Agent AI for High-Throughput Polymer Informatics: From Property Prediction to Generative Design Across Synthetic and Bio-Polymers

We present an integrated multiagent AI ecosystem for polymer discovery that unifies high-throughput materials workflows, artificial intelligence, and computational modeling within a single Polymer Research Lifecycle (PRL) pipeline. The system orchestrates specialized agents powered by state-of-the-art large language models (DeepSeek-V2 and DeepSeek-Coder) to retrieve and reason over scientific resources, invoke external tools, execute domain-specific code, and perform metacognitive self-assessment for robust end-to-end task execution. We demonstrate three practical capabilities: a high-fidelity polymer property prediction and generative design pipeline, a fully automated multimodal workflow for biopolymer structure characterization, and a metacognitive agent framework that can monitor performance and improve execution strategies over time. On a held-out test set of 1,251 polymers, our PolyGNN agent achieves strong predictive accuracy, reaching R2 = 0.89 for glass-transition temperature (Tg ), R2 = 0.82 for tensile strength, R2 = 0.75 for elongation, and R2 = 0.91 for density. The framework also provides uncertainty estimates via multiagent consensus and scales with linear complexity to at least 10,000 polymers, enabling high-throughput screening at low computational cost. For a representative workload, the system completes inference in 16.3 s using about 2 GB of memory and 0.1 GPU hours, at an estimated cost of about $0.08. On a dedicated Tg benchmark, our approach attains R2 = 0.78, outperforming strong baselines including single-LLM prediction (R2 = 0.67), group-contribution methods (R2 = 0.71), and ChemCrow (R2 = 0.66). We further demonstrate metacognitive control in a polystyrene case study, where the system not only produces domain-level scientific outputs but continually monitors and optimizes its own behavior through tactical, strategic, and meta-strategic self-assessment.

💡 Research Summary

The paper introduces a unified, multi‑agent artificial intelligence ecosystem designed to accelerate polymer discovery by integrating high‑throughput materials workflows, large language models (LLMs), and advanced machine‑learning (ML) techniques into a single Polymer Research Lifecycle (PRL) pipeline. At its core is a DeepSeek‑V2 based Planner Agent that decomposes complex scientific tasks into subtasks and delegates them to a suite of specialized agents: a Molecular Modeling Agent (using RDKit and graph neural networks, GNNs, to predict properties such as glass‑transition temperature, density, tensile strength, and elongation), a Physics‑Informed Agent (incorporating physics‑informed neural networks, PINNs, to enforce mechanistic constraints), an Ensemble‑Learning Agent (aggregating predictions and quantifying uncertainty via model consensus), as well as auxiliary agents for literature research, structural characterization, safety assessment, synthesis planning, reporting, and execution monitoring.

The workflow begins with a polymer’s SMILES string, which is transformed into a molecular graph and fed to the PolyGNN model for property prediction. Predicted properties then serve as inputs to a generative design LLM that proposes new polymer structures and experimental plans aligned with user‑defined objectives (e.g., target Tg range, mechanical performance). A safety screening module filters out chemically hazardous or impractical candidates, and a synthesis agent suggests viable monomers, catalysts, and processing conditions. All outputs are synthesized into a coherent report by a Reporting Agent, while an Execution Agent ensures fault‑tolerant, reproducible operation across the entire pipeline.

Empirical evaluation on a held‑out test set of 1,251 polymers demonstrates strong predictive performance: R² = 0.89 for Tg, 0.91 for density, 0.82 for tensile strength, and 0.75 for elongation. On a dedicated Tg benchmark of 50 polymers, the system outperforms strong baselines—single‑LLM (R² = 0.67), group‑contribution methods (R² = 0.71), and ChemCrow (R² = 0.66)—achieving R² = 0.78, a success rate of 0.76, and an efficiency score of 0.37. Computationally, inference for a typical workload completes in 16.3 seconds, consumes ~2 GB of RAM, 0.1 GPU‑hours, and costs roughly $0.08, scaling linearly to at least 10,000 polymers with a five‑fold speed‑up under parallel execution.

A distinctive contribution is the incorporation of meta‑cognitive self‑assessment: each agent logs performance metrics that are analyzed at tactical, strategic, and meta‑strategic levels, enabling the system to adapt its execution strategy over time. This capability is showcased in a polystyrene case study where the framework autonomously generates scientific outputs, monitors its own behavior, and iteratively refines its approach. A parallel protein‑structure case study illustrates the framework’s extensibility to biopolymers.

Limitations identified include incomplete modeling of long‑range polymer chain interactions, dependence on the quality and bias of existing polymer databases, and the current lack of direct integration with laboratory automation hardware, which still requires human intervention for synthesis. Future work aims to incorporate multi‑scale physical models, tighter coupling with robotic synthesis platforms, and expanded, high‑quality datasets to realize a fully autonomous, end‑to‑end polymer research assistant.