Page 128 - Read Online

Page 20 of 31 Li et al. J. Mater. Inf. 2026, 6, 10

and high-throughput tasks needs to be optimized, which affects the overall efficiency in high-throughput
experimental scenarios. Finally, given increasingly complex material systems and multidimensional
performance indicators, the interpretability and decision transparency of agents remain crucial for achieving
end-to-end trusted automation.

Data analysis and conclusion summary
The vast amounts of data generated from materials synthesis experiments and characterization need to be
analyzed and summarized into scientific conclusions. In the analysis of materials experimental data, AI
agents integrate ML, multimodal data, and cross-scale knowledge reasoning, demonstrating
multidimensional capabilities in data integration, key factor identification, and process optimization
decision-making. They enable automated parsing of complex experimental data, summarization of
conclusions, and optimization of experimental strategies, significantly enhancing the depth and efficiency of
data interpretation.

In multimodal data analysis and conclusion summary, agents can integrate experimental and simulation data
across sources and scales, revealing underlying mechanisms and universal laws that are difficult to discover
by traditional methods. For example, in SSE research, agents revealed a universal principle that hydrides
containing neutral molecules (e.g., NH ) have higher ionic conductivity by integrating experimental and
3
computational data from 158 metal hydrides. Through meta-dynamics simulations, a novel “two-step
migration” mechanism was discovered, providing theoretical guidance for the design of divalent
electrolytes . SciLink automatically transforms experimental data into verifiable scientific claims, quantifies
[122]
data novelty through literature comparison, and independently proposes subsequent experimental plans,
thereby enabling the systematic capture of “serendipitous discoveries” in materials data . Additionally, the
[123]
PROTeomics Exploration and Understanding System (PROTEUS) agent automatically analyzed 12
proteomics datasets and generated 191 scientific hypotheses . The CellVoyager agent is an AI system that
[124]
autonomously explores single-cell RNA sequencing datasets guided by prior user analyses. It parses historical
data, identifies new mechanisms, and translates expert intuition into verifiable analytical pathways, offering
valuable guidance for integrated multi-omics analysis in materials science [125] . Overall, these studies
demonstrate how agents not only integrate multimodal and cross-scale data but also reveal potential
mechanisms and generate novel hypotheses, effectively transforming raw experimental data into
interpretable and actionable scientific knowledge.

In experimental design and optimization, agents achieve dynamic regulation and multi-objective
optimization of synthetic pathways and material formulations through RL, generation models, and real-time
data fusion. Taking the development of ultra-high-performance concrete materials as an example, an agent
based on RL analyzes historical experimental data to construct a Markov decision process model, enabling
dynamic adjustment of material ratios and process parameters. Through six rounds of experimental
iterations, it increased the compressive strength from 55 to 221 MPa while reducing costs by over 25%. It also
identified the key strategy of balancing strength and cost through the regulation of cement and silica fume
content [126] . In the synthesis of MOFs, MOFsyn is a framework leveraging LLMs to guide the efficient
synthesis and performance optimization of MOFs. It integrates literature knowledge with real-time
experimental data to analyze relationships among synthesis conditions, characterization results, and catalytic
performance. It identified Ni content and H adsorption strength as the core factors influencing catalytic
0
2
activity and proposed a stepwise reduction strategy to guide the optimization of MOF synthesis. This enabled
Ni@UiO-66(Ce)-R2T1 to achieve 100% conversion and selectivity in olefin hydrogenation reactions . For
[127]
the design of transition metal complexes, LLM-EO, which integrates LLMs into evolutionary optimization,
combines generative models with multi-objective optimization, revealing the regulatory effects of strong
electron-withdrawing ligands on performance using only 20 experimental samples. In a chemical space of