A Numerically Aware GPT-Lite Method for Aluminum Alloy Composition Design
Xingci Hou1,2, Yingli Liu1,2,*, Jiancheng Yin3,*, Haihe Zhou1,2
1 Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, China
2 Yunnan Key Laboratory of Computer Technologies Application, Kunming University of Science and Technology, Kunming, China
3 Faculty of Materials Science and Engineering, Kunming University of Science and Technology, Kunming, China
* Corresponding Author: Yingli Liu. Email: 
; Jiancheng Yin. Email: 
Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.089174
Received 15 July 2026; Accepted 08 September 2026; Published online 20 September 2026
Abstract
AI for Science is extending materials modeling from property prediction toward inverse design. This study proposes Generative Pre-trained Transformer-lite (GPT-lite), a numerically aware framework for target-property-driven aluminum alloy composition design. Using a processed dataset containing 3553 composition–property records, 25 elemental variables, and three mechanical properties—elongation, tensile strength, and yield strength—the framework models forward property prediction and inverse composition generation as structured sequence tasks. EnhancedTokenizer is introduced to reduce numerical fragmentation, while a prefix-aware regression head provides continuous-property supervision to the lightweight Transformer backbone. In inverse design, generated compositions are subjected to strict parsing, compositional clipping and normalization, and unified evaluation by an external multilayer perceptron (MLP) surrogate. Under fixed data splits, 100 held-out targets, and the same candidate budget, GPT-lite achieves a mean best maximum relative error of 0.029679 and success rates of 0.85, 0.97, and 0.98 at relative-error thresholds of 0.05, 0.08, and 0.10, respectively, outperforming Non-dominated Sorting Genetic Algorithm II (NSGA-II), conditional variational autoencoder (CVAE), and conditional Wasserstein generative adversarial network with gradient penalty (cWGAN-GP) under the unified comparison protocol. Independent zero-shot validation on 40 external aluminum-alloy compositions yields R
2 values of 0.420, 0.748, and 0.740 for elongation, tensile strength, and yield strength. These results support the feasibility of combining numerically aware sequence modeling, continuous-property supervision, and constraint-aware evaluation for lightweight aluminum alloy inverse design.
Keywords
Aluminum alloy; composition design; inverse design; GPT-lite; numerically aware encoding; property prediction
Open Access
Login
Register
Submit a Paper
Propose a Special lssue
Download PDF
Downloads
Citation Tools
