AIonopedia: an LLM agent orchestrating multimodal learning for ionic liquid discovery¶

🕒 Published (v1): 2025-11-14 12:53 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

AIonopedia is a ReAct-driven LLM agent (powered by GPT-5) that orchestrates six specialized tools and a multimodal foundation model to automate ionic liquid (IL) property prediction, molecular screening, and design. It addresses the triple bottleneck of scarce labeled data, fragmented workflows, and poor model accuracy in IL research. Wet-lab validation on out-of-distribution systems confirms practical efficacy beyond offline benchmarks.

Problem¶

IL discovery faces: (1) scarce labeled experimental data relative to the vast cation-anion combinatorial space; (2) no effective fusion of inherently multimodal molecular data (graphs, SMILES, descriptors); (3) fragmented, non-automated chemical pipelines that require expert-in-the-loop intervention at each stage.

Method¶

Agent layer: A GPT-5 planner follows the ReAct paradigm—iterating Thought → Action → Observation—to invoke six tools: web searcher (Serper + Semantic Scholar APIs with RAG fallback), PubChem searcher (name-to-SMILES with LLM disambiguation), SMILES canonicalizer (RDKit), data processor (Python interpreter), property predictor, and molecule searcher.

Property predictor (core module): A dual-tower model with an LLM-based language encoder (LoRA fine-tuned) and a Graph Transformer encoder. Cross-attention decoders fuse the two views. Training uses a two-stage strategy: (1) modality alignment on 2.8 million synthetic IL virtual-system samples with momentum contrastive learning (InfoNCE loss), graph-text grounding (GTG, autoregressive cross-entropy), and graph-text matching (GTM, binary classification) losses plus a supervised MSE/cross-entropy loss on 21-dimensional pseudo-label descriptors; (2) fine-tuning on ~100K curated labeled samples replacing alignment FFNs with task-specific regression heads.

Molecule searcher: Hierarchical pipeline—exhaustive search over the property dataset for top-\(K\) candidates, followed by beam search in known IL databases and PubChem guided by Tanimoto similarity—transforms IL design into a heuristic search rather than generation, avoiding chemically unrealistic outputs.

Key Contributions¶

First LLM-based agent explicitly tailored to IL discovery (AIonopedia)
First LLM-augmented multimodal domain foundation model for ILs (two-stage alignment + fine-tuning)
Largest curated IL solute-solvent interaction dataset (~100K labeled samples, ~80 ionic solvents, ~150 solutes for \(\Delta G\) properties; ~6,000 bulk property systems with ~1,200 cations)
2.8M synthetic virtual-system alignment dataset constructed via combinatorial sampling over known IL components
Hierarchical molecular screening pipeline with Tanimoto-similarity beam search, validated in wet lab

Results¶

GPT-5 significantly outperformed all other tested web-enabled LLMs on ion abbreviation-to-canonical-SMILES conversion (50-ion benchmark evaluated on canonical SMILES match, structural match, and full name match)
Full AIonopedia system outperformed standalone GPT-5 at varying tool-integration levels on the data extraction task (Fig. 3B; specific \(\Delta\) not available in provided text)
Property predictor achieves superior performance across all property datasets relative to baselines (specifics in truncated results)
Modality alignment pretraining improves performance across all metrics versus no-alignment ablation (Section 4.4)
Agent demonstrated OOD generalization on wet-lab IL screening tasks outside the training distribution

Limitations¶

Generative molecular design abandoned; the combinatorial/search approach is bounded by known IL chemical space in databases
Fewer than 10,000 experimentally known ILs constrains alignment-stage diversity despite synthetic augmentation
Property predictor covers only two property classes (solute-solvent interactions and bulk properties); other IL-relevant properties not addressed
Full quantitative results unavailable in the provided text excerpt (results section truncated)
Dependence on GPT-5 API introduces cost, latency, and reproducibility concerns for large-scale screening

Relevance to Harnesses / Meta-Harnesses¶

AIonopedia is a domain-specific meta-harness: a ReAct planner that dynamically composes heterogeneous submodules—web retrieval, database lookup, ML inference, code execution, and similarity search—into a closed-loop pipeline, which is the defining architectural pattern of scientific harnesses. The two-stage training regime (alignment then fine-tuning) is itself a meta-level orchestration of learning pipelines, showing how harness design extends into model construction. The web-searcher's RAG fallback and the PubChem searcher's LLM disambiguation layer illustrate graceful-degradation strategies relevant to any robust harness. Wet-lab validation closes the loop from automated pipeline to physical experiment, demonstrating harness design at the cyber-physical boundary.