Inverse Materials Design by Large Language Model-Assisted Generative Framework

TL;DR

This paper presents AlloyGAN, a closed-loop inverse materials-design framework that combines Large Language Model–assisted text mining with a conditional GAN to generate alloy compositions meeting target properties. By constructing a richly described dataset from literature and thermodynamic descriptors, the CGAN can produce property-driven compositions and iteratively validate them experimentally, achieving predictions within $8\%$ of measured values. Downstream tasks demonstrate high accuracy in GFA classification ($\approx$0.90 F1) and strong regression performance ($R^2$ up to $0.80$ for key properties), underscoring the framework's versatility. The approach highlights the value of integrating LLM-derived knowledge with generative modeling and experimental feedback to accelerate autonomous materials discovery, particularly for multi-component metallic glasses.

Abstract

Deep generative models hold great promise for inverse materials design, yet their efficiency and accuracy remain constrained by data scarcity and model architecture. Here, we introduce AlloyGAN, a closed-loop framework that integrates Large Language Model (LLM)-assisted text mining with Conditional Generative Adversarial Networks (CGANs) to enhance data diversity and improve inverse design. Taking alloy discovery as a case study, AlloyGAN systematically refines material candidates through iterative screening and experimental validation. For metallic glasses, the framework predicts thermodynamic properties with discrepancies of less than 8% from experiments, demonstrating its robustness. By bridging generative AI with domain knowledge and validation workflows, AlloyGAN offers a scalable approach to accelerate the discovery of materials with tailored properties, paving the way for broader applications in materials science.

Figures (8)

• Figure 1: Overview of the AlloyGAN framework. The AlloyGAN framework integrates four key components for targeted composition design and discovering new alloys with specific properties. (I) LLM-Assisted Text Mining: We utilize the Uni-SMART LLM [15] and chemical prompts to extract relevant data from alloy literature. (II) Data Preprocessing: Extracted data is converted into chemical descriptors, with additional descriptor data generated automatically through mathematical formulas to enhance diversity and comprehensiveness. (III) Model Construction: A generative deep learning framework is developed to establish mappings between alloy compositions and properties. The framework employs Conditional Generative Adversarial Networks to generate new alloys that meet specific property requirements. (IV) Experimental Feedback Loop and Downstream Tasks: Properties of generated alloys are validated through experiments. Verified data is incorporated into the alloy dataset, creating a feedback loop that iteratively enhances model performance and dataset robustness. The framework also supports downstream applications, such as materials classification and property prediction.
• Figure 2: Expanded dataset comparison and distribution. (a) The dataset collected through LLM-based data mining exhibits broader coverage and enhanced diversity compared to the previous dataset zhou2023generative, highlighting the effectiveness of our approach in capturing a wider range of alloy compositions and properties. (b) Histogram of thermodynamic parameters ($T_l$, $T_x$, $T_g$) for alloys from zhou2023generative, along with the alloy data collected by LLM in this work.
• Figure 3: Comparison of GAN-generated and original alloy compositions across different alloy categories. (a)-(d): Two-dimensional PCA plots illustrating the distributions of GAN-generated and original samples for Cu-, Fe-, Zr-, and Ti-based alloys, respectively. The green solid lines represent the distribution of the original data collected through LLM-assisted data mining, while the blue dashed lines show the distribution of the data generated by the GAN model. The close alignment of the distributions highlights the ability of the GAN model to effectively capture the characteristic features of the original dataset.
• Figure 4: Inverse alloy design using conditional GAN. Two-dimensional PCA plots compare the generated samples from AlloyGAN with the original ones across four alloy categories. Data points in the latent space are color-coded to represent their respective categories. Confidence ellipses in corresponding colors highlight the distribution of each category, showing the alignment between generated and original data.
• Figure 5: Performance evaluation of machine-learning models. (a) Training stability comparison between GAN and CGAN models over 1250 generator updates, assessed using Jensen-Shannon (JS) divergence and Kullback-Leibler (KL) divergence. (b) Wasserstein distance ($d_w$) comparison among GAN, GAN+, and CGAN models. GAN uses only alloy composition as input, while GAN+ incorporates both composition and property features as inputs. (c) Classification accuracy comparison between our workflow and prior approaches, showing a significant improvement in categorizing alloys based on specific criteria. (d) Regression accuracy comparison for the properties $T_{rg}$ and $\gamma$.