Prediction and inversion methods for gas atomization metal powder size distribution inspired by fine tuning the large-scale pre-trained language models

Abstract

Powder size distribution plays an important role in mechanical properties of metallic components produced through additive manufacturing. This study introduces the novel prediction and inversion methods on powder size distribution during gas atomization. Both methods are inspired by fine tuning the large-scale pre-trained language models, which are comprised of Backpropagation, Genetic Algorithm, and Low Rank Adaptation (BpGA-LoRA) algorithm. In both methods, BpGA serves as the pre-trained model, while LoRA is employed as the fine-tuning model. The results demonstrate their superior regression performance on powder size distribution during gas atomization. In particular, when tackling complex regression challenges, such as simultaneously considering both effects of process parameters and nozzle geometry, the BpGA-LoRA-based predication method outperforms the traditional BpGA-based method, achieving a relative error reduction of 9.10 % for the training set and 6.56 % for the validation set. Additionally, the Earth Mover's Distance (EMD) is significantly decreased by 0.0177 for the training set and 0.0239 for the validation set. Similarly, the BpGA-LoRA model, when extended to the inversion method on powder size distribution affected by both effects of process parameters and nozzle geometry, delivers a relative error reduction of 6.12 % for the training set and 4.65 % for the validation set, as compared with the traditional BpGA model. Therefore, the proposed BpGA-LoRA model would not only enhance the development of powder production in additive manufacturing, but also open new avenues for predictive and inverse solutions in various industrial sectors.