CatDT — A Multi-Agent Catalysis Digital Twin
Project Description
CatDT (Catalysis Digital Twin) is an AI-driven, multi-agent platform for
heterogeneous catalysis simulation. Built on the CAMEL agent framework, it
orchestrates 8 LLM agents that call deterministic scientific tools to turn a
single bulk crystal structure and a natural-language reaction description into
reconstructed surfaces, validated reaction pathways, activation barriers, and
microkinetic rates.
The core idea: **the agent reasons and coordinates (LLM), while the tool does
the science (Python).** A key innovation is the Agent4/5 "Propose–Validate"
loop, where a pathway-design agent and a hard-gate validation agent iterate
(up to 10 rounds) to produce physically valid NEB endpoints before any
expensive barrier calculation runs. A memory system (Memento casebank) lets
the agents improve from past successes and failures without model fine-tuning.
This project aims to (1) extend CatDT's liquid-solid (electrochemical) track,
and (2) build a benchmark suite to evaluate the multi-agent system's pathway
design quality across reaction families (HER / OER / CO2RR / NRR).
heterogeneous catalysis simulation. Built on the CAMEL agent framework, it
orchestrates 8 LLM agents that call deterministic scientific tools to turn a
single bulk crystal structure and a natural-language reaction description into
reconstructed surfaces, validated reaction pathways, activation barriers, and
microkinetic rates.
The core idea: **the agent reasons and coordinates (LLM), while the tool does
the science (Python).** A key innovation is the Agent4/5 "Propose–Validate"
loop, where a pathway-design agent and a hard-gate validation agent iterate
(up to 10 rounds) to produce physically valid NEB endpoints before any
expensive barrier calculation runs. A memory system (Memento casebank) lets
the agents improve from past successes and failures without model fine-tuning.
This project aims to (1) extend CatDT's liquid-solid (electrochemical) track,
and (2) build a benchmark suite to evaluate the multi-agent system's pathway
design quality across reaction families (HER / OER / CO2RR / NRR).
Supervisor
CHENG, Lixue
Quota
2
Course type
UROP1000
UROP1100
UROP2100
UROP3100
UROP3200
UROP4100
Applicant's Roles
1. Study and adapt the CatDT codebase (`camel_agents/`, `core/`) to understand
the 8-agent workflow and the 28 CAMEL FunctionTools.
2. Help implement the liquid-solid electrochemical path: Pourbaix-mode
VSSR-MC reconstruction and constant-potential barrier evaluation.
3. Build a benchmark suite of catalytic reactions and run comparative
experiments measuring pathway validity and barrier convergence rates.
4. Analyze results, document findings, and draft a workshop paper.
the 8-agent workflow and the 28 CAMEL FunctionTools.
2. Help implement the liquid-solid electrochemical path: Pourbaix-mode
VSSR-MC reconstruction and constant-potential barrier evaluation.
3. Build a benchmark suite of catalytic reactions and run comparative
experiments measuring pathway validity and barrier convergence rates.
4. Analyze results, document findings, and draft a workshop paper.
Applicant's Learning Objectives
1. Design and run a tool-augmented, multi-agent scientific system.
2. Connect chemistry knowledge (surface science, microkinetics) with AI tools.
3. Gain hands-on experience with LLM agents, ML force fields (UMA / CHGNet),
and NEB / microkinetic modeling.
4. Develop and evaluate a domain-specific benchmark for AI4Science.
2. Connect chemistry knowledge (surface science, microkinetics) with AI tools.
3. Gain hands-on experience with LLM agents, ML force fields (UMA / CHGNet),
and NEB / microkinetic modeling.
4. Develop and evaluate a domain-specific benchmark for AI4Science.
Complexity of the project
Challenging