Page 119 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 11 of 31

Figure 4. The method of generating material structures by agents. (A) The overview of the MatAgent framework [64] ; (B) Pipeline for
training and generation using the MolGPT model [66] . Figure 4A is reproduced from “Accelerated Inorganic Materials Design with
Generative AI Agents”, arXiv:2504.00741, under CC BY 4.0 license [64] . Figure 4B is reproduced from “MolGPT: Molecular Generation
Using a Transformer-Decoder Model”, J. Chem. Inf. Model. 2022, 62, 9, 2064-2076, with permission from American Chemical Society [66] .
LLM: Large language model; GNN: graph neural network; RDKiT: RDKit toolkit; LogP: the logarithm of the partition coefficient; TPSA:
topological polar surface area; SAS: synthetic accessibility score; QED: quantitative estimate of drug-likeness; MolGPT: molecule
generative pre-trained transformer.

As the generative capabilities of LLMs have improved, new approaches have emerged that position LLMs as
the core engine for material structure generation, supplemented by specialized tools for verification and
optimization. These methods directly leverage the generative capabilities of LLMs by treating material
structures, such as SMILES (Simplified Molecular Input Line Entry System) strings or text representations of
Crystallographic Information File (CIF), as a specialized language. For example, MolGPT (Molecular
Generation Using a Transformer-Decoder Model), inspired by text-generation models, uses a
Transformer-decoder architecture to train on and generate molecular SMILES sequences [Figure 4B].
[66]
Further research indicates that pretrained LLMs without fine-tuning on materials data inherently possess the
ability to generate stable crystal structures. For instance, MatLLMSearch combines a pretrained LLM with
evolutionary search to achieve high success rates in crystal structure generation and performance
optimization .
[67]

More advanced systems have developed complex MAS centered around LLMs. MT-Mol introduces a