Technical Summary: OSCAgent

Problem Statement

The discovery of high-performance organic solar cell (OSCs) materials is currently hindered by a reliance on inefficient trial-and-error strategies. Traditional molecular generation methods, such as fragment recombination, variational autoencoders (VAEs), and genetic algorithms, often struggle to generate realistic candidates. These approaches typically focus on optimizing known molecular backbones and fail to effectively integrate domain-specific chemical knowledge, resulting in molecules that are either chemically invalid, synthetically inaccessible, or lack the diversity required to explore new chemical spaces. Furthermore, the scarcity of experimentally verified high-performance OSC molecules makes it difficult to train robust, generalizable generative models from scratch.

Methodology: OSCAgent Framework

To address these challenges, the authors propose OSCAgent, a multi-agent framework driven by Large Language Models (LLMs) that unifies retrieval-augmented design, molecular generation, and systematic evaluation into a continuously improving pipeline. The framework operates without requiring additional human intervention and consists of three collaborative agents:

1. The Planner: Acts as the strategic coordinator. It employs a retrieval-augmented strategy to ground its reasoning. It retrieves knowledge from two sources:

   • A Reference Database of experimentally confirmed, high-performance OSC molecules from literature (selected for structural diversity via a K-center greedy algorithm).
   • A Candidate Database of top-performing candidates dynamically updated from previous design cycles.
     The Planner synthesizes this information to formulate research plans and guide the Generator, while also reflecting on evaluation feedback to refine future strategies.

2. The Generator: Operates under the Planner's guidance to propose novel OSC acceptor molecules. It translates research directions into concrete molecular designs, emphasizing chemical plausibility and synthetic accessibility rather than unconstrained generation.

3. The Experimenter: Performs a comprehensive, multi-dimensional evaluation of candidate molecules using cheminformatics and machine learning tools. It generates structured reports to feed back to the Planner. The evaluation covers three key dimensions:

   • Efficiency: Predicted Power Conversion Efficiency (PCE).
   • Synthetic Accessibility: Estimated feasibility of synthesis.
   • Physical Feasibility: Consistency of electronic properties (HOMO/LUMO levels) with empirical ranges.

Core Technical Components

• Uncertainty-Aware Multi-Modal PCE Prediction: The Experimenter utilizes a specialized predictor trained on both the Lopez computational dataset (pretraining) and the Sun experimental dataset (fine-tuning).
  • Architecture: It combines graph and SMILES embeddings with handcrafted molecular fingerprints (Morgan, MACCS) processed via a Mixture-of-Experts (MoE) encoder.
  • Uncertainty Quantification: To address experimental noise, the model treats PCE as a random variable, outputting both a mean ($\mu$) and variance ($\sigma^2$). It is trained using a loss function that minimizes prediction error while calibrating predictive variance.
  • Auxiliary Tasks: The model includes an auxiliary task for predicting LUMO energy levels to strengthen the representation of photovoltaic-relevant features.
• Retrieval-Augmented Design: The system integrates static literature knowledge with dynamic feedback. The scoring function for retrieving top candidates balances predicted PCE, synthetic accessibility penalties, and rewards for orbital energies within empirically observed ranges.

Key Contributions

1. OSCAgent Framework: The introduction of an LLM-driven multi-agent system specifically for OSC acceptor discovery that autonomously generates higher-quality candidate molecules.
2. Enhanced Evaluation: The development of a multimodal PCE predictor with uncertainty quantification, which improves prediction accuracy and provides reliable, probabilistic feedback for molecular design.
3. Retrieval-Augmented Strategy: A novel design approach that integrates diverse, validated molecules with dynamically updated top candidates, enabling an exploration strategy that effectively balances novelty with chemical feasibility.

Experimental Results

The authors evaluated OSCAgent against traditional molecular generation methods (BRICS, VAE, SMILES-GA, Graph-GA) and LLM-based baselines (BioT5, Few-shot + Direct Reasoning).

• Performance Metrics: OSCAgent outperformed all baselines across seven metrics, including diversity (Uniqueness, Novelty), molecular quality (Validity, Average PCE), and distributional similarity to real high-performance molecules (measured via Sinkhorn–Wasserstein distance over multiple fingerprints).
• Key Findings:
  • Validity & Efficiency: OSCAgent achieved the highest validity (0.705) and the highest average predicted PCE (14.59%), significantly surpassing traditional methods which often generated chemically infeasible candidates.
  • Ablation Studies: Removing the Retrieval-Augmented Strategy or the Experimenter resulted in clear degradation across multiple metrics. Specifically, without the Experimenter, candidate validity and predicted efficiency dropped noticeably, highlighting the necessity of systematic chemical feedback.
  • Case Studies: Representative molecules generated by OSCAgent preserved hallmark structural motifs of high-performance OSCs (e.g., extended $\pi$-conjugation, electron-withdrawing cyano groups) while introducing novel variations. These candidates achieved predicted efficiencies approaching 18% with synthetic accessibility scores (SAscore) consistently below 8.0.
• PCE Prediction: The proposed PCE predictor achieved an $R^2$ of 0.713 and an MAE of 1.686 on the experimental dataset, outperforming recent neural architectures (e.g., RingFormer, GRIT) and traditional fingerprint-based models. Ablation confirmed that uncertainty quantification, pretraining, and handcrafted features were all critical to this performance.

Significance and Claims

The paper claims that OSCAgent demonstrates the effectiveness of integrating LLMs with domain-specific tools and retrieval strategies to accelerate scientific discovery. By creating a closed-loop workflow that continuously adapts and improves, the framework enables the efficient discovery of OSC acceptor molecules that are chemically valid, synthetically accessible, and possess superior predicted photovoltaic performance compared to existing methods. The authors position this work as a step toward reducing the reliance on trial-and-error in materials science, offering a scalable approach for exploring broader regions of chemical space. Future work is noted as extending the framework to other functional materials and integrating wet-lab feedback for tighter design-validation coupling.