Page 121 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 13 of 31

models. For instance, MatSciAgent establishes a multi-agent framework where a primary agent parses
natural language user queries and distributes tasks, such as materials data retrieval, crystal structure
generation, continuum simulations, and molecular dynamics calculations, to specialized agents. These agents
leverage databases such as the Materials Project and customized code to complete the computations .
[71]
Similarly, platforms such as Aitomia (an AI-powered platform for atomistic and quantum chemical
simulations) , ChemGraph (an agentic framework for computational chemistry workflows) , El Agente
[73]
[72]
(an autonomous agent for quantum chemistry) , MatAgent (an intelligent agent specifically designed for
[74]
predicting material properties) , and AutoSolvateWeb (a chatbot-assisted computational platform for
[75]
quantum chemistry studies) use chatbot interfaces and AI agents to decompose complex tasks into
[76]
subtasks. This approach simplifies the setup of complex simulations, including excited-state calculations,
thermochemical analysis, and quantum mechanics [Figure 5A]. These platforms also proactively guide users
through the setup, execution, and analysis of atomic-scale simulations, effectively integrating diverse
methods such as first-principles calculations, molecular dynamics, and ML-based potential functions,
thereby facilitating the broader adoption of atomic simulation technologies. To address the need for complex
input files and execution codes in computational tools, MechAgents enable autonomous coding,
[77]
correction, and execution of finite element analysis scripts through the collaboration of planning, coding,
and execution agents, solving elastic mechanics problems under different boundary conditions. The
language-to-simulation (Lang2Sim) framework achieves accurate conversion from text descriptions to
[78]
executable code for computational tools by decomposing simulation engines into tool functions and
input-output pairs. Molecular Dynamics Agent (MDAgent) automatically generates and optimizes
[79]
LAMMPS (Large-scale Atomic/Molecular Massively Parallel Simulator) molecular dynamics simulation code
by fine-tuning LLMs, reducing the average task time by approximately 42.22%. This stands in sharp contrast
to the traditional manual preparation of input files and script writing. Such research, by building “natural
language-computational tools” interfaces and decomposing tasks, makes core computational tools such as
density functional theory (DFT) and molecular dynamics more accessible and executable for materials
property calculation. These efforts advance the development of “language-driven simulation” and
significantly improve computational efficiency and accessibility.

Beyond automating traditional tools, by using ML models to calculate material properties, AI agents can
rapidly conduct high-throughput screening and multi-objective optimization in the materials space, breaking
through the efficiency limitations of traditional screening based on computational tools and achieving
intelligent search and design in the material space. For instance, agents adopting an exploration-utilization
strategy combined with ML models autonomously conduct stability screening in binary/ternary compound
spaces, successfully discovering hundreds of novel stable materials [80,81] . This is impractical through sequential
DFT calculations. QMLMaterial (a quantum machine learning software for material design and discovery)
efficiently identifies globally optimal structures of atomic clusters, surface adsorption systems, and related
materials using active learning algorithms and performance-evaluation ML models . For specific material
[82]
systems, MAS demonstrates highly efficient screening capabilities. In the field of alloys, LLM-driven MAS
collaborates with graph neural networks (GNN) to autonomously explore the design space of NbMoTa alloys
and predict macroscopic mechanical strength, accelerating the discovery of advanced alloys [Figure 5B]. In
[83]
zeolite, after the LLM agent proposes a candidate structure, atomic simulation evaluations provide iterative
feedback, efficiently generating novel molecules that exhibit higher binding affinity . In metamaterials,
[84]
agents automatically generate performance prediction models based on datasets and model architectures,
and combine developed generative models to achieve automated inverse design of metamaterials .
[85]
Moreover, ChemReasoner combines LLM-generated catalyst hypotheses with atomic structure evaluation of
GNNs, using metrics such as adsorption energy to guide the search direction, thereby improving the
efficiency of catalyst design . SciToolAgent drives hundreds of scientific tools through knowledge graphs
[86]
and demonstrates powerful automation capabilities in MOF screening scenarios [87,88] . MDLab enables