LacMaterial: Large Language Models as Analogical Chemists for Materials Discovery

Hongyu GuoarXiv 2025 · arXiv · OpenReview

TL;DR

This paper introduces "LacMaterial," a novel approach that leverages Large Language Models (LLMs) to perform analogical reasoning for materials discovery, specifically for solid-state electrolytes. By guiding LLMs with explicit cross-domain and in-domain analogies, the method generates promising new material candidates that outperform traditional prompting techniques and expand the search space beyond conventional approaches.

What Problem Does This Paper Address?

Scientific discovery often relies on analogical reasoning – drawing parallels between different systems to gain new insights. However, human experts are often limited by their domain-specific knowledge and biases, leading them to overlook "far" analogies that could be highly fruitful. This is particularly true in materials science, where discovering novel materials with specific properties (like those for next-generation batteries) is crucial but challenging. While LLMs have shown potential for analogical reasoning, their application in scientific discovery, especially in generating novel hypotheses, remains underexplored and often fragile, heavily dependent on prompt engineering. This paper aims to bridge this gap by developing structured analogical frameworks to make LLMs more robust and effective hypothesis generators for materials discovery.

Methods

The core of this work lies in two distinct analogical reasoning strategies powered by LLMs:

  1. Cross-Domain Analogical Reasoning: This approach leverages LLMs' vast training data to draw analogies from unrelated domains to inspire new materials. The LLM is prompted to identify abstract structural similarities between a target domain (e.g., solid electrolytes) and other domains (e.g., data-center networks, airports). It then generates analogy-guided exemplars and structural motifs from these analogous systems. These exemplars are not meant for direct copying but as conceptual strategies to inspire novel material designs. The LLM is also instructed to generate specific prompts for downstream material generation based on these analogies.
  2. In-Domain Analogical Reasoning: This strategy focuses on leveraging existing knowledge within the materials science domain, particularly when dealing with limited labeled data (e.g., a small set of known solid-state electrolytes with their conductivity values). The LLM is used to construct interpretable analogical rules and templates from this limited data. These templates capture key relationships between material composition, structure, and properties, enabling the LLM to generate new candidates by interpolating or extrapolating from known examples.

Both strategies emphasize generating candidates that are absent from existing databases and literature, ensuring novelty. Charge neutrality is a critical constraint enforced throughout the generation process.

Datasets

  • For Cross-Domain Analogies: The LLM's internal training data, which is vast and cross-domain, is implicitly used. Specific retrieval examples were drawn from domains like cathode and anode design.
  • For In-Domain Analogies:
    • A dataset of 600 expert-validated solid-state electrolyte entries with conductivity (σ) values.
  • An additional 578 entries from the Inorganic Crystal Structure Database (ICSD), resulting in a combined set of 1,178 inorganic, solid-state, lithium-containing compounds.
  • A corpus of 324 peer-reviewed papers and 21 review articles related to Li-ion conductivity.

The paper does not introduce a new benchmark dataset but rather uses existing data to develop and validate its analogical reasoning framework.

Key Insights and Ideas

  • LLMs as "Analogical Chemists": The paper positions LLMs not just as pattern matchers but as interpretable, expert-like hypothesis generators capable of leveraging analogy for scientific innovation.
  • Beyond Surface-Level Similarity: The core idea is to move beyond superficial similarities in material composition or structure and focus on deeper, relational analogies to uncover novel design principles.
  • Cross-Domain Analogies Unlock Novelty: Analogies from seemingly unrelated fields (like data-center networks) can provide powerful conceptual frameworks for designing materials with desirable properties (e.g., robust Li-ion transport).
  • In-Domain Analogies for Data-Efficient Discovery: When data is scarce, LLMs can distill analogical rules from a small set of examples to guide targeted exploration within a specific domain.
  • Structured Analogical Frameworks Enhance Robustness: Explicitly defining analogical reasoning strategies makes LLM outputs more systematic and reliable compared to naive prompting.

Key Results

  • Novel Material Candidates: The LLM successfully generated novel solid-state electrolyte candidates for cubic LLZO, many of which were absent from established databases like ICSD and Materials Project.
  • Outperformance of Naïve Prompting: Candidates generated using the analogical reasoning strategies (both cross-domain and in-domain) consistently received higher "vote counts" (indicating consensus among simulated generation rounds) and were more diverse than those generated by a vanilla, non-analogical prompt.
  • Top Candidate Example: The top-ranked candidate from the cross-domain analogy was Li7.00La2.50Nd0.50Zr1.40Hf0.60O12, inspired by the "data-center backbone" analogy, which aims to ensure Li-ion transport redundancy.
  • Computational Validation: All five top candidates from the cross-domain analogy showed negative total energies, indicating thermodynamic favorability through MACE-MP surrogate calculations.

Limitations and Future Work

  • Computational Expense of Property Prediction: A significant challenge remains in computationally evaluating critical properties like ionic conductivity, which often requires expensive molecular dynamics simulations.
  • Expanding Analogy Templates: The authors suggest expanding the diversity of analogy templates to cover a wider range of scientific reasoning types.
  • Integrating with Surrogate Models: Integrating LLM-driven generation with efficient surrogate property models is proposed as a key step towards enabling closed-loop, language-guided materials pipelines.
  • Rapid Validation: The ultimate goal is to couple LLM reasoning with rapid experimental validation for faster materials discovery.

Summary updated

Abstract

Analogical reasoning, the transfer of relational structures across contexts (e.g., planet is to sun as electron is to nucleus), is fundamental to scientific discovery. Yet human insight is often constrained by domain expertise and surface-level biases, limiting access to deeper, structure-driven analogies both within and across disciplines. Large language models (LLMs), trained on vast cross-domain data, present a promising yet underexplored tool for analogical reasoning in science. Here, we demonstrate that LLMs can generate novel battery materials by (1) retrieving cross-domain analogs and analogy-guided exemplars to steer exploration beyond conventional dopant substitutions, and (2) constructing in-domain analogical templates from few labeled examples to guide targeted exploitation. These explicit analogical reasoning strategies yield candidates outside established compositional spaces and outperform standard prompting baselines. Our findings position LLMs as interpretable, expert-like hypothesis generators that leverage analogy-driven generalization for scientific innovation.
BibTeX
@misc{guo2025lacmaterial,
  author = {Hongyu Guo},
  title = {{LacMaterial: Large Language Models as Analogical Chemists for Materials Discovery}},
  year = {2025},
  eprint = {2510.22312},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.22312},
}
Figure