PaperlayerView exact revision on arXiv

Source-linked AI summary

Accelerated Inorganic Materials Design with Generative AI Agents

Izumi Takahara, Teruyasu Mizoguchi, Bang Liu

arXiv:2504.00741v1cond-mat.mtrl-sci

TL;DR

Inorganic materials discovery needs scalable ways to explore target properties beyond expert-intensive or exhaustive computational searches. MatAgent combines LLM reasoning, external tools, structure generation, and property evaluation in iterative feedback; the authors report improved compositional uniqueness and novelty while maintaining validity and comparably high success rates.

  • Problem

    Traditional computational discovery can require expert modeling or exhaustive candidate enumeration, while generative and LLM-based approaches have limitations in target precision, interpretability, and compositional exploration.

    Source: Introduction · pp. 1–2

  • Method

    MatAgent uses an LLM with external memory and materials tools, structure estimation, and property evaluation to iteratively refine compositions toward target properties.

    Source: Introduction · p. 2, MatAgent framework · p. 2, LLM as agent · p. 4

  • Results

    Tool-assisted Planning and Proposition significantly improved compositional uniqueness and novelty while maintaining comparably high success rates.

    Source: Iterative materials generation guided… · p. 8, Discussion · p. 12

  • Takeaways & Limitations

    External resources support broader materials-space exploration and practical applicability beyond the LLM's inherent limitations.

    Source: Discussion · p. 12

    The framework guides generation toward a single target property, formation energy, rather than simultaneous optimization of multiple properties.

    Source: Discussion · p. 12

Abstract

from arXiv · show

Designing inorganic crystalline materials with tailored properties is critical to technological innovation, yet current generative computational methods often struggle to efficiently explore desired targets with sufficient interpretability. Here, we present MatAgent, a generative approach for inorganic materials discovery that harnesses the powerful reasoning capabilities of large language models (LLMs). By combining a diffusion-based generative model for crystal structure estimation with a predictive model for property evaluation, MatAgent uses iterative, feedback-driven guidance to steer material exploration precisely toward user-defined targets. Integrated with external cognitive tools-including short-term memory, long-term memory, the periodic table, and a comprehensive materials knowledge base-MatAgent emulates human expert reasoning to vastly expand the accessible compositional space. Our results demonstrate that MatAgent robustly directs exploration toward desired properties while consistently achieving high compositional validity, uniqueness, and material novelty. This framework thus provides a highly interpretable, practical, and versatile AI-driven solution to accelerate the discovery and design of next-generation inorganic materials.

Introduction

Materials discovery must explore broader spaces more efficiently and with less reliance on human experts. MatAgent combines LLM reasoning, external tools, structure estimation, and property evaluation to support interpretable, iterative exploration toward target properties.

  • Traditional computational discovery often depends on expert modeling or exhaustive candidate enumeration, limiting efficiency and scalability across broader materials spaces.Source: Introduction · p. 1
  • Generative models have been applied to crystalline-material generation, but typical single-step approaches may not precisely meet target properties.Source: Introduction · p. 1
  • Target-aware methods often optimize in latent spaces that obscure reasoning, while LLM methods remain compositionally limited or require target-specific fine-tuning.The paper calls for scalable frameworks that incorporate external knowledge and enable interpretable, iterative refinement toward target properties.Source: Introduction · pp. 1–2
  • MatAgent centers a general-purpose LLM and combines four external tools with structure estimation and property prediction for interpretable, iterative refinement.The tools are short-term memory, long-term memory, the periodic table, and a materials knowledge base.Source: Introduction · p. 2

Results

The results assess MatAgent’s iterative success and compositional novelty across runs, including the effects of tool-assisted planning and natural-language initialization. They also examine how natural-language elemental constraints shape generated compositions and their adherence to specified restrictions.

Iterative materials generation guided by target property

  • GPT-4o favored the knowledge base and long-term memory for −3.8 eV/atom targets but added periodic-table use at −1.6 eV/atom, while o3-mini relied heavily on the knowledge base for the harder target.Both LLMs increasingly used long-term memory in later iterations; o3-mini also increased short-term memory use, while GPT-4o reduced its early reliance on the knowledge base.Source: Iterative materials generation guided… · p. 5, Figure 3 · p. 5
  • Success rates increased over iterations; GPT-4o with tools outperformed its no-tools variant especially at −3.8 and −2.5 eV/atom, while o3-mini without tools outperformed its tool-assisted variant.MatterGen typically had higher success rates early on. For o3-mini without tools, high success at −3.8 eV/atom was likely driven by repeatedly proposing known stable compounds such as HfO2.Source: Figure 4 · p. 6, Iterative materials generation guided… · pp. 6–7

Compositional metrics

  • 0.20 (GPT-4o w/o tools) to 0.60 (GPT-4o w/ tools) and 0.0 (o3-mini w/o tools) to 0.55 (o3-mini w/ tools) were the respective V.U.N. scores.o3-mini w/o tools had compositional validity 1.0 but novelty 0.0; about 60% of its candidates were HfO2. Novelty comparisons with MatterGen may be affected by differences in training data.Source: Figure 5 · p. 7, Iterative materials generation guided… · pp. 7–8

Elemental distributions

  • Tool-assisted proposals covered more elements: GPT-4o shifted beyond Ti and O toward F, Ca, and Al, while o3-mini added F, Th, and Y beyond its usual Hf and O.Without tools, the LLMs often proposed well-known compositions with high success but limited novelty; tool assistance improved uniqueness and novelty while maintaining comparably high success rates.Source: Iterative materials generation guided… · p. 8, Figure 6 · p. 9

Individual tool contributions

  • Individual-tool experiments suggested that the knowledge base supports unique compositions with high success, long-term memory improves uniqueness and novelty, and the periodic table aids exploration of novel elemental spaces.The reported evidence summarizes Table 1 qualitatively and provides no cell values in the supplied passages.Source: Iterative materials generation guided… · p. 8, Table 1 · p. 8

Natural language integration to initialization

  • Retriever initialization began closer to the −3.8 eV/atom target than random initialization and achieved higher overall success, while its slightly lower novelty left V.U.N. scores comparable across initialization methods.LLM initialization also began closer to the target than random initialization.Source: Natural language integration to… · pp. 8, 10, Figure 7 · p. 10

Constrained generation with natural language

  • The framework was tested with natural-language constraints excluding Pb, Hg, Cd, and Cr; excluding actinides; or restricting proposals to non-metal elements.The supplied passages describe these constraint conditions but do not report their results.Source: Constrained generation with natural… · p. 10, Figure 8 · p. 11

Results

  • Without constraints, generated compositions spanned diverse element classes; imposed constraints excluded Pb, Hg, Cd, and Cr, nearly eliminated actinides, or favored non-metals.Under Constraint 3, compositions were predominantly composed of O, B, Si, P, F, N, and Ge.Source: Constrained generation with natural… · p. 11
  • Constraint 1 raised compositions excluding Pb, Hg, Cd, and Cr from 0.96 to 1.00; Constraint 2 raised actinide-excluding compositions from 0.97 to 0.99; Constraint 3 raised exclusively non-metal compositions from 0.00 to 0.83.Figure 8 reports elemental distributions aggregated over 10 independent runs of 15 iterations each.Source: Constrained generation with natural… · p. 11

Discussion

MatAgent combines interpretable LLM composition proposals with external tools and structure and property models to iteratively guide inorganic materials toward target properties. The study reports gains in uniqueness and novelty with validity maintained, while identifying single-property scope and model dependence as limitations.

Discussion

  • MatAgent uses explicit LLM reasoning, four external tools, a diffusion-based Structure Estimator, and a GNN-based Property Evaluator to refine compositions toward user-specified targets.The evaluator and estimator provide feedback that guides subsequent proposals.Source: Discussion · pp. 11–12
  • External tools increased the uniqueness and novelty of proposed compositions while maintaining compositional validity in generation targeting formation energies.The study also reports that Retriever-based initialization was effective and that LLMs support flexible constrained generation.Source: Discussion · p. 12
  • The current framework targets only one property—formation energy—and multi-objective design would require evaluators for diverse properties and more complex feedback.The authors also note that the framework does not assess the correctness of the LLM’s reasoning and depends strongly on the Structure Estimator and Property Evaluator.Source: Discussion · p. 12
  • The authors identify integrating synthesizability considerations and synthesis-related knowledge as a future direction for practical autonomous materials discovery.They propose that LLM text-based reasoning could support assessment of proposed materials’ practical synthesizability.Source: Discussion · p. 12

Methods

The methods combine four retrieval tools with structure and property models, and compare random, LLM-based, and Retriever-based composition initialization.

Tools

  • Short-term memory retrieves the three most recent proposals, their reasoning, and feedback to help identify trends and revise potentially redundant proposals.The number of recent proposals retrieved in this study is three.Source: Short-term memory retrieval · p. 12
  • Long-term memory retrieves the top three past proposals that successfully generated compositions close to the target property, together with their reasoning.These retrieved cases provide examples of successful proposals near the target.Source: Long-term memory retrieval · p. 12
  • The periodic-table tool retrieves same-group elements, while knowledge-base retrieval ranks five composition-similar entries and selects the three the LLM judges most relevant.The knowledge base contains 10,000 entries built from composition pairs and GPT-4o-generated explanations of property changes.Source: Periodic table retrieval · p. 12, Knowledge base retrieval · p. 12
  • The 48,323-structure MP-60 dataset trains the Structure Estimator and Property Evaluator; DiffCSP generates five candidate structures per reduced composition, subject to a 34-atom limit.The dataset uses structures with up to 60 atoms per cell and energy above hull no greater than 0.02 eV/atom, then splits them 6:2:2.Source: Dataset · p. 13, Structure Estimator · p. 13
  • The Property Evaluator uses iComFormer to predict formation energy per atom and selects the candidate structure with the lowest predicted value as most stable.It also returns format-correction feedback when a proposed composition is inadequate.Source: Property Evaluator · p. 13

Initialization of composition

  • Random initialization samples initial compositions from MP-20.Source: Random sampling · p. 13
  • LLM-based initialization asks the LLM to propose an initial composition directly in response to a natural-language formation-energy target.Source: LLM-based sampling · p. 13
  • Retriever-based initialization uses contrastive learning with T5 to select initial compositions relevant to natural-language descriptions of desired properties.It encodes property descriptions and corresponding chemical formulas from MP-20.Source: Retriever-based sampling · p. 13

Author contributions statement

The authors report distinct roles in study design, framework development, evaluation, supervision, and manuscript preparation.

Author contributions statement

  • I. T. designed and developed the study framework, ran experiments, analyzed results, and drafted the manuscript; T. M. assisted evaluation and editing, while B. L. supervised and edited.All authors reviewed the manuscript.Source: Author contributions statement · p. 15

A Prompt templates

The appendix provides prompt templates for planning, tool-conditioned composition proposals, feedback, a no-tool proposal, and knowledge-base observation generation.

A.1 Prompt template for the Planning stage

  • The Planning prompt supplies the target and current situation, including prior tool choices and their resulting property changes.Source: Figure 9 · p. 16, §A.1 · p. 16

A.3 Prompt template for feedback generation

  • The feedback prompt uses a property-value variable; invalid or unreadable composition formats receive instructions to resubmit correctly.Examples include invalid element symbols and fractional subscripts.Source: §A.3 · p. 16, Figure 14 · p. 19

A.5 Prompt template for observation generation

  • The knowledge-base observation prompt provides source and target compositions, formula units, and property values for generating transition explanations.Source: §A.5 · p. 16, Figure 16 · p. 19

A.2 Prompt templates for the Proposition stage

  • The short-term-memory Proposition prompt includes recent proposals, their reasoning, and Property Evaluator feedback.Source: §A.2 · p. 16, Figure 10 · p. 17
  • The long-term-memory Proposition prompt includes successful past proposals near the target, their reasoning, and evaluator feedback.Source: §A.2 · p. 16, Figure 11 · p. 17
  • The periodic-table Proposition prompt supplies same-group elements; the example maps Li to H, Na, K, Rb, Cs and O to S, Se, Te.Source: §A.2 · p. 16, Figure 12 · p. 18
  • The knowledge-base Proposition prompt supplies insights about property changes under related composition modifications and explanations for those changes.Source: §A.2 · p. 16, Figure 13 · p. 18

A.5 Prompt template for observation generation

  • The no-tool template prompts the LLM to propose a material composition using target, composition, and property-value variables.Source: Figure 15 · p. 19

A.4 Prompt template for comparative experiments

Figure 15’s prompt template lets the LLM revise a proposed composition using feedback on its previous proposal.

A.4 Prompt template for comparative experiments

  • The LLM receives feedback on its previous composition and returns a revised proposal.Figure 15 shows the prompt template used without tool-assisted Planning and Proposition.Source: §A.4 · p. 16

B Data distribution

Figure 17 presents the formation-energy distribution in the MP-60 dataset, with dashed lines marking selected quantiles.

B Data distribution

  • Figure 17 marks MP-60 formation-energy quantiles at 1.0% (−3.8 eV/atom), 2.5% (−3.5 eV/atom), 10% (−3.0 eV/atom), 20% (−2.5 eV/atom), and 40% (−1.6 eV/atom).The quantiles are calculated by sorting formation energies in ascending order.Source: B Data distribution · p. 20, Figure 17 · p. 20

C Reasoning examples

The reasoning analysis finds tool-associated patterns in Planning, but less distinct clustering in Proposition. The authors also report chemically relevant proposal rationales while noting that their correctness is not guaranteed.

C Reasoning examples

  • Planning-stage reasoning forms distinct clusters that correspond to selected tools, and knowledge-base reasoning sometimes reflects prior successful use.The analysis uses SciBERT embeddings and t-SNE; Figure 18 colors points by k-means cluster, and the knowledge-base pattern also appears for o3-mini.Source: C Reasoning examples · pp. 20–21, Figure 18 · p. 20
  • Proposition-stage reasoning does not form clearly separable t-SNE clusters, unlike Planning-stage tool-selection reasoning.The authors suggest this difference may reflect composition proposal’s flexibility compared with choosing among four fixed tools.Source: C Reasoning examples · p. 21, Figure 19 · p. 21
  • Proposition reasoning uses relevant chemical keywords, including bonding types, electronegativity, and atomic affinity, but its physical and chemical correctness is not guaranteed.The authors suggest refining these explanations and improving their reliability could make them useful insights for human experts.Source: C Reasoning examples · p. 21

Related papers

Loading 2504.00741v1…