Page 101 - Read Online

Page 8 of 15 Wang et al. J. Mater. Inf. 2026, 6, 9

the task was reasonably interpreted and decomposed into several sub-tasks by Matty. It then autonomously
constructed a modular simulation pipeline and executed each computational step, as illustrated in Figure 3.
The resulting workflow demonstrates a higher level of complexity than that in Task 1 due to the involvement
of a broader suite of computational tools, including the CVAE generative model, RDKit for conformer
analysis, Gaussian for quantum chemical optimization, and Turbomole combined with COSMO-RS for
solvation energy calculations. Notably, Matty correctly identified the problem as a molecular design task and
selected a CVAE as the appropriate generative model based on user-specified constraints such as molecular
weight (MW), HOMO, and LUMO energy levels. The CVAE model was trained using data from the QM9
dataset and an in-house database, enabling the generation of chemically valid candidate molecules
represented as Simplified Molecular Input Line Entry System (SMILES) strings. Each candidate molecule
underwent 3D conformer generation using RDKit, followed by geometry optimization at the B3LYP/6-31G
(2df, p) [87,88] level of theory using Gaussian. Molecules that satisfied the target electronic criteria were then
subjected to solubility evaluation. The Agent selected Turbomole with the COSMO-RS solvation model to
compute solvation free energy (G ) in a battery-relevant solvent environment [ethylene carbonate
sol
(EC)/dimethyl carbonate (DMC) = 1:1.5 molar ratio], using the B-P-D3BJ/def-TZVP level of theory [89-91] .
Finally, Matty compiled all relevant properties, including MW, HOMO, LUMO, and G , into structured
sol
JSON-format outputs. These newly evaluated structure-property pairs were subsequently reintegrated into
the in-house database to iteratively refine the CVAE model. This feedback-driven loop not only improved
the quality of candidate generation but also enhanced Matty’s generalizability over time, establishing a
scalable and adaptive framework for data-efficient molecular discovery.

To assess the reliability of our AI-Agent in the workflow, we performed a series of internal validation tests
across multiple task categories. For each task, the Agent was prompted with natural language instructions
and tasked with constructing and executing the corresponding workflow. In our trials, the Agent achieved a
high success rate, with all tested cases executing without observable errors. This high level of robustness is
attributed to the MCP tool design with clearly defined functions and application scenarios, structured tool
metadata, and the strong reasoning capabilities of the underlying LLM. As a representative demonstration,
we conducted three conditional molecular generation tasks, each defined by distinct electronic and structural
constraints. These case studies underscore Matty’s ability to flexibly translate user-defined objectives into
coherent, multi-step workflows, showcasing both technical reliability and domain adaptability. Matty designs
the molecules to satisfy the following performance specifications: high solubility in EC/DMC (1:1.5 mole
ratio), a MW of approximately 100 Da, HOMO energy above -6.0 eV, and LUMO energy below 0.0 eV. For
demonstration purposes, Matty is configured to perform only a single screening iteration. Two more design
cases are further considered, with details provided in the Supporting Information (SI). A set of SMILES with
the properties were outputs, and the results were summarized in Supplementary Table 1.

In Case 1, Matty generated 28 molecules, 17 of which matched entries in the QM9 dataset, while 11 were
novel structures not previously recorded. Case 2 produced 25 molecules, including 17 known
electrolyte-related species and 8 new candidates. Similarly, in Case 3, Matty proposed 28 molecules, of which
18 were present in QM9 and 10 were newly generated. As shown in Figure 4A-C and Supplementary Table 2,
although the generated molecules do not exactly match the specified targets, their property distributions
exhibit distinct shifts across the three design scenarios, demonstrating Matty’s ability to conditionally
modulate generation behavior based on task-specific inputs. Representative molecules from each case,
selected for proximity to the mean values of the target properties, are visualized in Figure 4D.

Furthermore, solubility-related metrics such as G of the new candidates were computed, as reported in
sol
Supplementary Table 3. Importantly, Matty not only completed the inverse design task based on electronic
and structural criteria but also recognized the electrolyte-relevant context and extended the workflow by