Physics-Embedded Multi-Task Neural Surrogate for Predicting and Optimizing Incremental Sheet Forming of AZ31B Magnesium Alloy
Karthikeyan.G, Sasirekha SP, Melbin J Reena, Benilin Leebha.B, Reshma R, Mekala N
National Institute of Technology Tiruchirappalli Karpagam Academy of Higher Education Christ University
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
AZ31B magnesium alloy ISF (Incremental Sheet Forming) requires reliable deformation prediction from simulation despite the strongly nonlinear effect of a few controlling process parameters. The present work develops a physics-embedded neural surrogate model using exclusively the real CNC ISF experiments and with three input variables available: spindle speed, feed rate and pitch. The surrogate is developed to anticipate a constrained though crucial forming responses: that is, the depth of forming, wall-thickness reduction and major strain which leads to a lightweight application-specific model. Physics-based regularization on the thinning behavior of the membrane and on the coupling between strain and pitch is incorporated in the learning phase to ensure physically reasonable results, also for modest datasets. The experimental verification results in good predictive capability, with R2of 0.958 for forming depth, R2= 0.944 for thickness reduction, and R2= 0.937 for major strain. MAEs of geometric responses are smaller than 0.28mm. Compared with a traditional ANN, the proposed model delivers decrescents of non-physical deviations by 33% and improvements in the prediction stability against untested parameter combination by 15-20%. Since the trained surrogate is differentiable, gradient-based parameter optimization can be used to select settings which yield an increase in achievable forming depth of 6-9% for values that satisfy safe thinning limits. The future work will consist on the ongoing development of the tool to account for temperature phenomena like in situ quench and tangent effects, tool shape variables, and active-learning loops to enhance predictivity power.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
工程Metal Forming Simulation Techniques
Laser and Thermal Forming Techniques · Machine Learning in Materials Science