Harjit Pal Singh
Department of Physics, The University of the West Indies, St. Augustine, Trinidad and Tobago.

DOI https://doi.org/10.33889/PMSL.2026.5.1.012

  ;

Abstract

These days, Machine Learning (ML) is widely used in material science research due to various advancements in algorithmic architecture. ML algorithms can be used for forward and inverse design prediction of the materials. Based on material compositions and processing, inverse design material informatics can produce any specific material profile. However, this machine learning approach is limited due to lack of data and presence of imbalanced datasets. The problem of imbalanced data severely impacts the outcome due to biased machine learning models that favor majority classes and ignore the rare classes in the datasets. To address these issues, the work in this paper proposed a comprehensive deep learning framework for inverse prediction for aluminium (Al) alloys using post-processing conditions. The proposed framework utilized the generative manifold-guided variational autoencoder (VAE) along with the deep learning-based TabTransformer for classification tasks. This transformer-based attention-aware VAE-guided framework worked well enough to capture complex interactions among compositional and mechanical features of Al alloy samples. The performance evaluation results demonstrated that significant improvement was achieved through the proposed framework of VAE-guided TabTransformer. The effectiveness of the proposed work was highlighted with a mass increment in macro F1-scores, which increased from 0.49 to 0.89 when compared to the baseline framework. The improvement was not only due to the majority classes, but the outcomes of minority classes also contributed significantly. Further, the performance was supported by K-fold and statistical validation results. The proposed framework successfully addressed the complex data-imbalanced inverse design problem and proved to be a powerful, generalizable, and intelligent tool for material discovery and optimization.

Keywords- Inverse prediction, VAE, Transformer, Machine learning, Deep learning.

Citation

Singh, H. P (2026). VAE-Guided Transformer Based Deep Learning Framework for Inverse Prediction of Aluminium (Al) Alloy Post-Processing Conditions. Prabha Materials Science Letters, (1), 221-243. https://doi.org/10.33889/PMSL.2026.5.1.012.