ChatMOF: An Autonomous AI System for Predicting and Generating Metal-Organic Frameworks

Introduction

• Complex materials are difficult to represent in text, limiting how well LLMs can understand and process them.The introduction says material representations often fail to capture complete properties.Source: Introduction · p. 3
• Materials science has fewer dedicated databases and associated data, making suitable text-format training data scarce.The passage links this scarcity to the challenge of representing the data in a form suitable for LLMs.Source: Introduction · p. 3
• Earlier materials-science LLM efforts mainly extracted literature data and answered questions from it, leaving the material itself largely untouched.The introduction presents direct exploration of materials as an unrealized opportunity.Source: Introduction · p. 4
• ChatMOF is an autonomous AI system focused on MOFs that predicts properties from text inquiries and generates MOFs with specified properties.The introduction describes the latter capability as inverse design.Source: Introduction · pp. 4–5
• MOFs are used in many chemical applications because of their porosity, high surface area, and tunability.The cited passage gives these properties as reasons for their applications.Source: Introduction · p. 4
• ChatMOF is presented as a potential bridge between novice users and computational or machine-learning tools for developing materials.The introduction also suggests this could facilitate material development for applications.Source: Introduction · p. 4

Design for ChatMOF

• ChatMOF handles unstructured textual queries, retrieving material information and generating MOF structures to meet requested properties.The system also aims to predict properties and deduce synthesis techniques.Source: Design for ChatMOF · p. 6
• Its agent plans a response, selects a toolkit, observes the output, and evaluates whether it can answer or should plan again.ChatMOF comprises an agent, toolkits, and an evaluator.Source: Design for ChatMOF · pp. 6–7

Toolkits

• ChatMOF offers table-searcher, internet-searcher, predictor, generator, and other toolkits for retrieving, predicting, or generating material information.Other tools include calculators, file operations, and visualization aids.Source: Toolkits · p. 9
• For lookup, ChatMOF draws on CoREMOF and QMOF data, MOFkey building-block information, and DigiMOF synthesis conditions.Table-search accuracy depends on available pre-calculated values, and cannot serve queries for properties absent from the files.Source: Toolkits · pp. 9–10
• MOFTransformer predicts selected properties using fine-tuned models; it was pretrained on one million hypothetical MOFs and supports predictions across properties.ChatMOF selects a model suited to the query rather than computing properties directly; lookup predictions depend on pre-calculated values, while simulation is time- and resource-intensive.Source: Toolkits · p. 12
• Atom-by-atom inverse design remains elusive for complex MOFs, and ChatMOF identifies genetic algorithms using selection, mutation, and crossover as suitable for LLM integration.The cited complexity includes many atoms, large void fractions, and complex topologies.Source: Toolkits · p. 15
• ChatMOF represents MOFs as textual genes encoding topology and building blocks, enabling an LLM to apply genetic algorithms.HKUST-1 is given as tbo+N17+N10.Source: Toolkits · p. 15
• For generation, ChatMOF plans an objective and loss function, selects matching parent genes, uses an LLM to create children, iterates toward the target, and returns the closest structure.Source: Toolkits · pp. 15–16
• Beyond LLM capabilities, ChatMOF provides file and Internet search and calculations, and uses ASE with planned Python code for material-structure operations.ASE supports atom manipulation, cell-information retrieval, and visualization.Source: Toolkits · p. 18

Evaluation

• GPT-4 outperformed GPT-3.5-turbo across tasks; GPT-3.5-turbo accuracy was 95% for search, 91% for prediction, and 77.8% for generation.The paper attributes GPT-4’s stronger accuracy to refined reasoning and comprehension, particularly during planning.Source: Evaluation · p. 20
• Search-task token-limit errors arise when code exceeds its allowance, while logic errors can reflect flawed strategies or code that fails to meet a requested item count.Token-limit errors often occur when listing all qualifying materials or handling comparisons; the paper gives a request for 10 items answered by extracting high values as an example of a flawed strategy.Source: Evaluation · pp. 20–21
• Prediction errors arise when interpreting machine-learning estimates or retrieving the correct table answer, and can follow unsuitable model choices or incompatible inputs.Source: Evaluation · p. 21
• Generation can fail when no parent genes satisfy a target range: ChatMOF expands the data, but errors may persist, and selecting the best generated MOF can also fail.The paper also notes token-limit and logic errors during extraction of the MOF best aligned with the objective.Source: Evaluation · p. 21
• With GPT-4, excluding token-limit-exceeded instances (4/100 search, 6/100 prediction, and 2/100 generation), accuracies were 96.9% for search, 95.7% for prediction, and 87.5% for generation.Source: Evaluation · p. 19, Figure 6 · p. 22

Inverse Design Validation

• The surface-area distribution shifted toward higher values across generations, with the average rising from 3,748 m2/g initially to 5,554 m2/g by the third generation.Source: Inverse Design Validation · p. 23, Figure 7 · p. 25
• For a target hydrogen uptake near 500 cm3/cm3, the initial distribution spanned 250–650 cm3/cm3, while the final generation formed a narrow peak at 500 cm3/cm3.Source: Inverse Design Validation · p. 23, Figure 7 · p. 25
• The generated rtl+N535+N234 structure had a predicted surface area of 6411.28 m2/g and an optimized accessible surface area of 7647.62 m2/g, ranking third in CoREMOF.Source: Inverse Design Validation · pp. 23–24, Figure 7 · p. 25
• The generated dia+N719+E186 structure had a hydrogen uptake of 499.998 cm3/cm3, matching the 500 cm3/cm3 target; geometric optimization and RASPA evaluation were also performed.Source: Inverse Design Validation · p. 24
• Token limits restrict parent and child structures to around 100, while resource and time constraints limit topologies and cycles; despite these limits, ChatMOF generates MOFs fitting the objective function.Source: Inverse Design Validation · p. 24

Collaborative Online Platforms

• Prediction depends on the number of pre-trained MOFTransformer weights, and the section proposes an online platform where users can post trained models for others to access.Source: Collaborative Online Platforms · p. 26
• The proposed platform would review new weights and download models when required data are available; shared models and pre-calculated data could support searches and collective improvements.Source: Collaborative Online Platforms · p. 26

Conclusion

• ChatMOF reported success rates of 96.9% for search, 95.7% for prediction, and 87.5% for structure generation.Source: Conclusion · p. 27

• The authors declare no conflicts to declare.Source: Conflicts of interest · p. 29

• Yeonghun Kang and Jihan Kim are listed as authors affiliated with KAIST's Department of Chemical and Biomolecular Engineering.Source: Supplementary information for · p. 33