Page 98 - Read Online

Wang et al. J. Mater. Inf. 2026, 6, 9 Page 5 of 15

Table 1. Main calculation tools supported by the Agent, with selected information included in tool metadata for ReAct
Typical system size Optimization time
Software/tool System type Accuracy rating Computational cost
(atoms) reference
Gaussian Molecular 50-200 High 0.5-2 h
Turbomole Molecular 50-500 Medium 0.3-1 h
XTB Molecular/Periodic 100-1,000 Medium 0.2-0.5 h
VASP Periodic 50-200 High > 2 h

3
LAMMPS Molecular/periodic 10 -10 6 Low 1 s
DimeNet++ Molecular 50-100 Medium 1 min
CGCNN Periodic crystal 100-1,000 Medium Seconds for inference
ReAct: Reasoning and Action; XTB: Extended Tight Binding; VASP: Vienna Ab Initio Simulation Package; LAMMPS: Large-scale Atomic/Molecular
Massively Parallel Simulator; CGCNN: Crystal Graph Convolutional Neural Networks.

(COSMO-RS) , Extended Tight Binding (XTB) , Vienna Ab Initio Simulation Package (VASP) ,
[51]
[52]
[50]
Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS) , Multiwfn , etc.
[53]
[54]
Machine learning tools: Gaussian process regression with Scikit-learn [55,56] , Multi-Objective Bayesian
Optimization with BoTorch , Conditional Variational Autoencoder (CVAE) , Graph Convolutional
[58]
[57]
Networks (GCN) , Graph Attention Networks (GAT) , DimeNet/DimeNet++ , Crystal Graph
[61]
[59]
[60]
Convolutional Neural Networks (CGCNN) with Torch and Torch-geometric .
[63]
[64]
[62]
Learning & optimization: Auto update database and ML models.
Given the relatively simple workflow of material computational design, we implemented Agent construction
using LangChain and the MCP Python Software Development Kit . The Agent adopted a hybrid
[65]
[66]
local/remote deployment architecture: the decision core and lightweight tools (LangChain, RDKit, etc.) were
hosted on a 32-core workstation, while computationally intensive tools (Gaussian, VASP, LAMMPS, etc.)
were deployed remotely and connected via the Server-Sent Events protocol.
To ensure that dynamically assembled workflows are both methodologically valid and logically coherent, we
implement a rigorous tool governance mechanism within Agent. Each tool is further encapsulated using the
MCP, which provides explicit metadata for describing tool capabilities and applicability. As shown in Table
1, metadata defines input/output formats, applicable systems and execution rules of tools. Metadata acts as a
prompt, allowing the Agent to understand the tool’s usage and assemble reasonable workflows. When a task
is initiated, the Agent combines the user’s objective and tools’ metadata via the MCP interface to perform
goal-oriented reasoning, translating natural language instructions into executable workflows. By combining
structured constraints with descriptive prompts, the system ensures that each generated workflow is not only
executable but also scientifically valid and reproducible.

RESULTS AND DISCUSSION
To evaluate the generality, robustness, and cross-domain applicability of our Agent, we present two
representative case studies encompassing both periodic and molecular material systems. The first case
focuses on a fully autonomous electronic structure calculation of monolayer transition metal dichalcogenides
(TMDs) , a prototypical 2D semiconductor, which highlights the Agent’s capacity to orchestrate
[67]
end-to-end density functional theory (DFT) workflows for periodic materials. The second case addresses the