Scholarlink

<p>Optimizing process parameters in laser-based Directed Energy Deposition (DED) remains challenging due to nonlinear interdependencies that influence geometric accuracy and mechanical performance. This study presents a machine learning–driven framework that integrates artificial neural network (ANN) surrogate modeling, evolutionary multi-objective optimization, and multi-method sensitivity analysis to enhance process design and decision-making in DED. The framework is experimentally validated using a 1&#xa0;kW laser system to deposit Inconel 718 onto stainless steel substrates. Two case studies are conducted: (i) optimization of bead geometry (width and dilution depth), and (ii) multi-layer cube fabrication targeting dimensional accuracy and surface hardness. The ANN model is trained on data generated via Design of Experiments, with its structure optimized using grid search to reduce prediction error and improve generalizability. Parameter influence is quantified through three independent sensitivity analysis methods, enabling transparency and interpretability in model behavior. Results show that laser power strongly influences bead width and build height, while nozzle speed and inter-bead spacing affect dilution and deposition uniformity. The Non-Dominated Sorting Genetic Algorithm II (NSGA-II) identifies Pareto-optimal parameter sets, achieving prediction errors below 5%, including dimensional error under 3% and hardness deviation within 3.5%. Unlike prior ANN–NSGA-II approaches, this study integrates a grid search–optimized ANN and applies three complementary sensitivity methods to enhance model interpretability and parameter insight. The modular and data-optimization and significantly reduces experimental workload, supports explainable optimization, and is generalizable to other additive processes. It provides a scalable and intelligent approach for process parameter design in laser-based additive manufacturing.</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

Machine learning–driven process parameter optimization in directed energy deposition: a multi-objective framework with experimental validation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月25日

查看全文
  • Parviz Kahhal,
  • Hong Seok Kim,
  • Mahmoud Moradi,
  • Sang-Hu Park

摘要

Optimizing process parameters in laser-based Directed Energy Deposition (DED) remains challenging due to nonlinear interdependencies that influence geometric accuracy and mechanical performance. This study presents a machine learning–driven framework that integrates artificial neural network (ANN) surrogate modeling, evolutionary multi-objective optimization, and multi-method sensitivity analysis to enhance process design and decision-making in DED. The framework is experimentally validated using a 1 kW laser system to deposit Inconel 718 onto stainless steel substrates. Two case studies are conducted: (i) optimization of bead geometry (width and dilution depth), and (ii) multi-layer cube fabrication targeting dimensional accuracy and surface hardness. The ANN model is trained on data generated via Design of Experiments, with its structure optimized using grid search to reduce prediction error and improve generalizability. Parameter influence is quantified through three independent sensitivity analysis methods, enabling transparency and interpretability in model behavior. Results show that laser power strongly influences bead width and build height, while nozzle speed and inter-bead spacing affect dilution and deposition uniformity. The Non-Dominated Sorting Genetic Algorithm II (NSGA-II) identifies Pareto-optimal parameter sets, achieving prediction errors below 5%, including dimensional error under 3% and hardness deviation within 3.5%. Unlike prior ANN–NSGA-II approaches, this study integrates a grid search–optimized ANN and applies three complementary sensitivity methods to enhance model interpretability and parameter insight. The modular and data-optimization and significantly reduces experimental workload, supports explainable optimization, and is generalizable to other additive processes. It provides a scalable and intelligent approach for process parameter design in laser-based additive manufacturing.