Scholarlink

<p>Amorphous alloy coatings show broad application prospects in surface protection, yet their corrosion rates are highly dependent on spray process parameters. Fragmented data, skewed corrosion rate distribution, and limited model interpretability in existing studies have severely hampered the research and development and industrial application of high-performance coatings. To address these challenges, this study established a comprehensive framework that balances accurate prediction and practical design. First, we developed an improved deep neural network (IDNN) model with two core innovations: (1) We proposed a kernel density estimation-based DenseLoss function to realize continuous adaptive weight assignment; this function fundamentally mitigates the adverse effects of imbalanced corrosion rate data spanning 6 orders of magnitude, and outperforms standard weighted loss functions; (2) We adopted the sparrow search algorithm (SSA) to optimize the model’s key hyperparameters. With its unique foraging-anti-predation mechanism, SSA effectively prevents the model from falling into local optima and enables the SSA-optimized model to achieve higher prediction accuracy than models optimized by other mainstream intelligent optimization algorithms. In addition, we applied a feature contribution analysis method to clarify the influence mechanisms of key factors on corrosion rate, which further improves the interpretability of the model’s prediction logic. The results show that the optimal <i>IDNN-SSA</i> model significantly outperforms existing methods and classic machine learning benchmarks in prediction accuracy, achieving an <i>R</i><sup>2</sup> of 0.956 on the training set and 0.945 on the independent test set. We identified core influencing factors and their optimal control ranges: reducing coating porosity and crystallization fraction, and regulating oxygen flow rate and powder delivery rate within the data-derived ranges, can significantly improve coating corrosion resistance. The SHAP-derived optimal ranges are statistical correlations requiring further single-variable experimental validation, yet theoretical predictions show the optimized scheme can reduce the corrosion rate of typical Fe-based amorphous coatings in 3.5 wt.% NaCl solution by over 40%. This study provides reliable data support, prediction tools, and optimization guidelines for corrosion control of amorphous alloy coatings, and is of great significance for promoting their large-scale engineering application.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable corrosion rate prediction and rational design of amorphous alloy coatings: a deep learning framework with DenseLoss and SHAP interpretability

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Zhuoxin Yang,
  • Zhilin Long,
  • Lidong You

摘要

Amorphous alloy coatings show broad application prospects in surface protection, yet their corrosion rates are highly dependent on spray process parameters. Fragmented data, skewed corrosion rate distribution, and limited model interpretability in existing studies have severely hampered the research and development and industrial application of high-performance coatings. To address these challenges, this study established a comprehensive framework that balances accurate prediction and practical design. First, we developed an improved deep neural network (IDNN) model with two core innovations: (1) We proposed a kernel density estimation-based DenseLoss function to realize continuous adaptive weight assignment; this function fundamentally mitigates the adverse effects of imbalanced corrosion rate data spanning 6 orders of magnitude, and outperforms standard weighted loss functions; (2) We adopted the sparrow search algorithm (SSA) to optimize the model’s key hyperparameters. With its unique foraging-anti-predation mechanism, SSA effectively prevents the model from falling into local optima and enables the SSA-optimized model to achieve higher prediction accuracy than models optimized by other mainstream intelligent optimization algorithms. In addition, we applied a feature contribution analysis method to clarify the influence mechanisms of key factors on corrosion rate, which further improves the interpretability of the model’s prediction logic. The results show that the optimal IDNN-SSA model significantly outperforms existing methods and classic machine learning benchmarks in prediction accuracy, achieving an R2 of 0.956 on the training set and 0.945 on the independent test set. We identified core influencing factors and their optimal control ranges: reducing coating porosity and crystallization fraction, and regulating oxygen flow rate and powder delivery rate within the data-derived ranges, can significantly improve coating corrosion resistance. The SHAP-derived optimal ranges are statistical correlations requiring further single-variable experimental validation, yet theoretical predictions show the optimized scheme can reduce the corrosion rate of typical Fe-based amorphous coatings in 3.5 wt.% NaCl solution by over 40%. This study provides reliable data support, prediction tools, and optimization guidelines for corrosion control of amorphous alloy coatings, and is of great significance for promoting their large-scale engineering application.