Interpretable Inverse Design of Metal-Organic Frameworks with Large Language Model Agents

This paper introduces LLM4MOF, a closed-loop multi-agent framework designed for the interpretable inverse design of metal-organic frameworks (MOFs). It addresses the challenge of searching a combinatorially vast space where property labels are expensive and traditional machine learning models often fail to provide chemical design insight, making it difficult to understand why specific structures succeed.

The LLM4MOF Framework

LLM4MOF operates as a system that separates chemical reasoning from structure construction. The workflow begins with a user-defined target query and proceeds through five distinct stages:

1. Hypothesis generation

2. Constraint translation

3. Candidate matching

4. Hypothesis testing

5. Feedback-driven refinement

The process is driven by two specialized agents. The first, a hypothesis generator, proposes building-block-level hypotheses involving metal nodes, linkers, pore geometry, and chemical functionality. The second agent, a constraint translator, maps these reasoning-based hypotheses into concrete search constraints linked to valid MOF candidates. This modularity ensures that the design intent remains interpretable while being grounded in automated computational modules.

Search and Evaluation Modes

The framework is validated through two complementary execution modes: database mode and discovery mode. In database mode, the agents are blind to the global property landscape, yet they successfully steer toward high-performing MOFs by retrieving properties from precomputed datasets. This mode tests the reasoning, translation, matching, and feedback loop without requiring new simulations.

In discovery mode, LLM4MOF performs inverse design of MOFs with live simulations. Rather than retrieving existing structures, it proposes and simulates new MOFs de novo. The system uses a deterministic assembler to build structures and evaluates them via automated GCMC simulations. This allows the framework to adapt geometry to requested conditions and arrive at design principles even in the absence of a precomputed reference.

Diagnostic Beams and Interpretability

A central feature of the framework is its ability to organize candidates into four diagnostic beams, which allows the system to determine which design axis drives performance. These beams include:

• Beam 1: The full-hypothesis beam

• Beam 2: The metal–linker chemistry beam

• Beam 3: The metal-only beam

• Beam 4: A random baseline beam

By comparing these beams, the Feedback Generator can identify whether performance improvements arise from geometry, chemistry, or metal choice. This enables the framework to provide interpretable design reasoning that is attributable to specific axes. For example, in adsorption tasks, the system may find that pore-geometry constraints are critical, whereas in electronic-structure tasks, chemical composition may carry more signal.

Efficiency and Performance

LLM4MOF demonstrates high efficiency by concentrating its search on top-performing structures within roughly 400 property evaluations. In discovery mode, it outperforms both random search and a genetic algorithm under an identical evaluation budget. The framework is highly cost-effective, with LLM inference costs estimated at approximately 1 per campaign. To manage the growing context of multi-turn conversations, the system utilizes a memory ledger that compresses earlier evidence into a fixed-size summary, ensuring that earlier observations do not dominate later reasoning.

1. Implementation of Multi-Agent Decoupled Reasoning and Execution

• Improvement: Architect AI systems to separate high-level semantic reasoning (Hypothesis Generation) from low-level parameter/constraint mapping (Constraint Translation). Instead of a single model attempting to generate complex structured data directly, use a specialized Reasoning Agent to produce interpretable, mid-level abstractions (e.g., JSON describing chemical properties) and a Translator Agent to map those abstractions into machine-executable code or search queries.

• System Capability: This prevents structural hallucinations and ensures that the AI's design intent is always compatible with downstream physical simulators or hardware controllers, significantly increasing the success rate of complex, multi-step generative tasks.

2. Integration of Diagnostic Attribution Beams for Explainable Optimization

• Improvement: Replace standard success/failure reward signals in optimization loops with a multi-channel Diagnostic Beam framework. When evaluating candidates, the system should simultaneously sample from multiple sub-hypotheses: the full hypothesis, a version with one variable held constant (e.g., chemistry only), and a baseline random sample.

• System Capability: This allows the AI to perform Attribution-Based Refinement. The system can mathematically determine whether a performance gain was driven by geometry, chemical composition, or specific elemental choices. It transforms the AI from a black-box optimizer into an interpretable researcher that understands why a specific configuration succeeded.

3. Deployment of Distilled Memory Ledgers for Long-Horizon Reasoning

• Improvement: Implement a hierarchical memory architecture that uses a Memory Ledger to compress historical iteration data into factual, non-prescriptive summaries rather than passing raw conversation history. This ledger should store only statistical frontiers (e.g., median values, descriptor envelopes, and trajectory trends) rather than full text logs.

• System Capability: This enables the AI to maintain stable reasoning over hundreds of iterations without suffering from contextual drowning, recency bias, or exponential increases in inference costs/token usage.

4. Hybrid Surrogate-Guided Discovery Loops

• Improvement: In high-cost evaluation environments (e.g., physical robotics or high-fidelity simulations), implement an asymmetric pre-filtering stage where a fast, low-fidelity surrogate model (like MOF2Zeo) is used to rank candidates specifically against the most expensive component of the current hypothesis.

• System Capability: This allows the system to perform intelligent resource allocation, concentrating expensive computational or physical resources only on candidates that satisfy the most critical geometric or structural constraints, thereby maximizing discovery efficiency under tight budgets.

5. Feature-Based Semantic Search over Library-Based Retrieval

• Improvement: Shift AI search strategies from Entity-ID retrieval (searching for specific known objects) to Feature-Space reasoning (searching for combinations of descriptors). The agent should reason over properties (e.g., short, rigid aromatic linkers) rather than specific database entries.

• System Capability: This enables the system to bridge heterogeneous datasets and perform de novo discovery. The AI can navigate between existing databases and entirely new generative spaces using a unified semantic language, allowing it to propose novel configurations that have never been explicitly recorded in its training data.

Inverse design of metal-organic frameworks (MOFs) requires navigating combinatorial spaces with costly property labels and opaque machine-learning models. We introduce LLM4MOF, a closed-loop multi-agent framework that converts a natural-language target into chemical hypotheses, constraints, diagnostic tests, and feedback. One agent proposes interpretable hypotheses over metal nodes, linkers, pore geometry, and functionality. Another converts them into constraints selecting MOFs defined by a node, linker, and topology. The Matchmaker forms four beams to attribute gains to geometry, chemistry, or metal choice: full hypothesis, chemistry, metal only, and random baseline. Blind to database landscapes, LLM4MOF enriches top performers across six adsorption, separation, and electronic-structure tasks within 400 evaluations. It also designs and live-simulates de novo MOFs spanning H2 storage, SF6 capture, and C2H6/C2H4 separation, deriving a distinct design rule for each objective. Under an identical nominal evaluation budget it consistently outperforms random search, Bayesian optimization, and genetic algorithms, and the outcome is insensitive to the language-model backend. All of this uses not a single property-labeled training structure, whereas generative alternatives train on thousands to hundreds of thousands. A new objective requires only a natural-language request: interpretable inverse design at a fraction of the data cost of existing approaches.