Page 99 - Read Online

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

generative design and multi-property evaluation of electrolyte additive molecules for lithium-ion batteries ,
[68]
demonstrating the integration of conditional generative models, quantum chemical calculations, and
solvation theory within a closed-loop inverse design pipeline. Collectively, these case studies illustrate the
Agent’s flexibility in addressing diverse materials design challenges, its ability to perform multi-step
reasoning and decision-making.

AI-Agent-enabled autonomous workflow for periodic TMDs: a representative example
To evaluate the capabilities of the proposed AI-Agent, Matty can be invoked with a simple instruction, for
example:

“Please calculate the electronic properties of MoS for me.”
2
Following that, Matty will comprehend this statement and break it down into the following task flow. As
illustrated in Figure 2, Matty designed a reasonable workflow for this task. It is apparent that Matty
successfully recognized that MoS is a periodic system, and selected VASP as the appropriate simulation tool
2
accordingly. Matty retrieved the crystal structure from the C2DB database and initiated a geometry
optimization using the PBE exchange-correlation functional with a plane-wave cutoff of 500 eV. It then
[69]
proceeds to a static electronic structure calculation, where the functional is upgraded to HSE06
(Heyd-Scuseria-Ernzerhof screened hybrid functional) to improve the accuracy of band gap prediction, a
[70]
common and validated practice for 2D semiconductors.

Once all computations are complete, Matty extracts key electronic properties such as the band structure,
band gap, and its nature (direct or indirect), and compiles the results in a structured JSON (JavaScript Object
Notation) format. As shown in Supplementary Figure 1, for monolayer MoS , the Agent predicts a direct
2
band gap of 2.34 eV, which aligns well with experimental data and previous theoretical studies [71,72] . This case
demonstrates Matty’s ability to autonomously construct and execute a complete DFT workflow, including
structure acquisition, parameter selection, convergence handling, hybrid-functional switching, and result
extraction, for periodic materials systems. The results not only validate the reliability of Matty’s
decision-making but also underscore its potential to automate routine yet technically demanding simulations
in materials science. In practice, the agent will also dynamically adjust its feedback based on computational
conditions. For instance, when convergence fails, VASP’s calculation parameters will be optimized according
to actual conditions.

Generative Agent-guided screening of electrolyte molecules
To further demonstrate the versatility of our Agent, we implemented a fully autonomous pipeline for the
inverse design and evaluation of molecules. While molecular systems account for only a subset of the broader
materials design landscape, including electrolyte additive [73,74] , energetic materials [75,76] , organic light-emitting
diode (OLED) emitters [77,78] , and deep eutectic solvents [79,80] , the design of functional molecules remains
pivotal in these domains. Importantly, the strategies developed for small molecules are readily extensible to
polymeric materials and formulation optimization, further broadening their relevance. As the demand for
high-performance batteries with increased voltage windows, enhanced cycle life, and superior safety
continues to grow, the performance requirements for electrolytes - and in particular, electrolyte additives -
have become increasingly stringent [81-85] . Rational design of such molecules often hinges on theoretical
evaluation of key electronic and physicochemical descriptors, including highest occupied molecular orbital
(HOMO) and lowest unoccupied molecular orbital (LUMO) energy levels (indicative of oxidative/reductive
stability), chemical hardness, and dipole moment [68,73] . These parameters offer a robust direction for assessing
redox reactivity, solubility, and functional performance .
[86]