Page 95 - Read Online

Page 2 of 15 Wang et al. J. Mater. Inf. 2026, 6, 9

INTRODUCTION
With the exponential growth of computational power and continuous advancements in algorithmic models,
computational materials science has evolved from a tool for interpreting experimental phenomena into a
core engine for goal-driven materials design . Leveraging quantum mechanics, molecular dynamics, and
[1,2]
multiscale simulation methods, researchers are now able to probe the intricate relationships of material
properties with atomic, molecular and crystal structures, as well as quantify the influence of processing
techniques on microstructural evolution, thus overcoming the limitations of traditional trial-and-error
approaches . The integration of machine learning (ML) with high-throughput computing has further
[3,4]
accelerated this transformation [5-7] . Intelligent predictive models that map structure-property-processing
relationships, combined with multiscale simulations (ab initio → molecular dynamics → phase-field
methods) and automated feedback loops, have enabled a paradigm shift from empirical screening to
on-demand design [8-10] . This synergistic co-design of performance, processing, and structure, which may also
integrate with manufacturing workflows, offers a seamless bridge from discovery to deployment, significantly
reducing barriers to industrial application.

High-throughput virtual screening strategies, driven by the principles of Materials Genome Engineering
(MGE), are already shortening development cycles and yielding disruptive materials for energy, electronics,
and beyond [11-18] . Yet, according to the standard MGE paradigm, the design of new materials still requires
researchers to manually integrate high-throughput simulation platforms, structure-property regression
models, and inverse design tools, often guided by expert intuition and system-specific heuristics [19,20] . This
process involves coordinating a diverse set of computational software, multiscale physical models, ML
algorithms, and databases [21,22] . Such integration imposes steep learning curves and considerable development
overhead, demanding interdisciplinary expertise and often resulting in rigid, non-transferable workflows that
are tailored to some narrow domains.

This raises a critical question: Can we construct an intelligent system capable of autonomously designing
workflow and dynamically composing tools to adapt to diverse material design tasks? Recent breakthroughs
in large language models (LLMs), particularly those with advanced reasoning capabilities, have made such
systems, artificial intelligence (AI) Agents. increasingly feasible [23-26] . The concept of an “Agent” was first
introduced in 1973 by Hewitt et al. through the Actor Model, which formalized key characteristics such as
autonomy, reactivity, and interactivity . While long dormant, this concept has been revitalized with the
[27]
emergence of LLMs in 2022 . Modern AI Agents leverage LLMs as cognitive engines, integrating modules
[28]
for perception, memory, decision-making, planning, execution, and learning into a cohesive, closed-loop
architecture: Understanding needs → Planning pathways → Invoking tools → Executing tasks → Evaluating
results → Dynamic optimization [29-31] . Such Agents are capable of autonomously achieving complex goals,
refining their workflows based on interaction history, and dynamically generating customized task
sequences.

Recent advances across multiple scientific disciplines underscore the transformative potential of LLM-based
AI Agents in automating complex, domain-specific workflows [32-38] . In spectral analysis, the SpeLL Agent (an
Agent for Natural Language-Driven Intelligent Spectral Modeling) enables end-to-end modeling of
near-infrared (NIR) spectra by generating domain-specific analytical scripts and using historical datasets to
guide algorithm selection . In computational biology, ProtChat integrates GPT-4 with protein-specific
[32]

language models to automate tasks such as protein property prediction and protein-drug interaction analysis,
significantly lowering the technical barrier for non-expert users . In the chemical sciences, the dZiner
[33]