ChatMOF

Overview

ChatMOF connects natural-language requests with computational workflows for metal-organic frameworks (MOFs). Its documented tasks are data retrieval, property prediction and structure generation, including generation directed by user-specified properties. The README describes a system built around GPT-4 and GPT-3.5 model options, rather than a standalone predictive model or a database alone. It can serve as a conversational interface for selecting and coordinating materials-research tools without requiring every request to be expressed as a formal structured query.

The workflow has three components: an agent interprets the request, develops a plan and chooses a toolkit; the selected tools produce task outputs; and an evaluator assembles a final response. Inputs are textual requests, while outputs depend on the chosen task: retrieved information, property predictions or generated MOF structures. The documented toolkit also includes a unit converter. The repository provides Python package configuration and a command-line entry point, so the resource is more than a paper-only proposal.

ChatMOF should nevertheless be approached as a research prototype whose deployment requires external services and task-specific dependencies. The online demonstration requires an OpenAI key and supports search; the README warns that prediction and generation generally do not work there beyond supplied examples. Local prediction and generation require module setup, and generation additionally requires GRIDAY. Linux is the recommended installation environment. Code availability and its MIT license do not establish the licensing or availability of linked datasets, external services or dependent model assets.

Key Features

• Interprets natural-language MOF requests and routes them through an agent, toolkit and evaluator workflow.
• Supports MOF data-retrieval tasks, with a search-focused online demonstration requiring an OpenAI key.
• Provides property-prediction workflows after local task-module setup.
• Supports MOF structure generation aimed at user-requested properties, with additional GRIDAY installation required.
• Includes a unit-conversion tool and command-line controls for model selection and temperature.

Use Cases

• Suggested evaluation: use natural-language search requests to retrieve MOF information, then check the returned information against its underlying records.
• Suggested evaluation: run locally configured property-prediction tasks and compare outputs with independently selected reference data.
• Suggested evaluation: explore property-directed MOF generation and assess resulting structures separately before treating them as scientifically validated candidates.

How to Use

1. Read the official README to choose between search, prediction and generation. Start with a narrow MOF question and specify what information or output you need.
2. For a search-only trial, open the online demo. It requires an OpenAI key; do not assume local prediction and generation capabilities are available in this interface.
3. For local work, follow the README installation instructions for pip install chatmof. The documented requirement is Python 3.9 or later, with Linux recommended. Consult its dependency troubleshooting notes if installation fails.
4. Configure OpenAI credentials privately using the documented environment variable. Run chatmof setup for prediction or generation, and chatmof install-griday for generation. Internet search additionally requires the documented Google credentials.
5. Start with chatmof run; use chatmof run --help to inspect options. Evaluate retrieved facts, predictions and generated structures independently. Consult the linked research paper for scientific context.