Scholarlink

<p>This research work utilizes machine learning, and taguchi method to optimize and predict the tensile performance of short carbon fiber reinforced polyamide (CF-PA) 3D printed by fused filament fabrication (FFF) technique. The process parameters play a major role in making novel material with help of FFF technology. In this research work, tensile samples are 3D printed based on different FFF process parameters, such as raster orientations, printer speed, and layer height as per ASTM D3039. While analysis of variance (ANOVA) confirms the statistical dominance of raster orientation (48.8% contribution), the framework’s novelty lies in leveraging a validated random forest model to uncover complex non-linear relationships and parameter interactions missed by traditional analysis. Methodologies including SHAP (Shapley Additive explanations) and partial dependence plots revealed a distinct non-linear effect of printer speed on tensile strength, peaking at 60&#xa0;mm/s. Furthermore, the model was used to generate a process interaction map, demonstrating its utility as a predictive tool for exploring the design space beyond the initial experimental runs. The final SVR and random forest models demonstrated the highest predictive precision with the lowest average mean squared error under rigorous leave-one-out cross-validation (LOOCV). This work validates a cost-effective methodology for achieving deep process understanding from sparse datasets, accelerating the adoption of FFF-printed composites. The combination producing maximum strength determined through S/N analysis is stated as a printer speed at 60&#xa0;mm/s with a 0<sup>0</sup> raster orientation having a layer height of 0.1&#xa0;mm. A strong predictive tool can be developed out of a comparison among four machine learning models (ML), namely XGBoost, support vector regression (SVR), linear regression (LR), and random forest (RF) using leave-one-out cross-validation (LOOCV) as the primary validation strategy with mean squared error (MSE) being the major performance metric. The SVR and random forest models return the maximum predictive precision with the reduces average MSE. SEM morphology studies on samples support results obtained quantitatively by relating good mechanical properties to better fibre-matrix bonding and low porosity seen at optimal settings. This study is validating a low-cost method of producing CF-PA components, that ensure high performance with definite mechanical properties, expediting their use in highly demanding engineering applications.</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

Taguchi-based experimental optimization coupled with explainable machine learning for predictive modelling of FFF-printed CF-PA composites

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Gokulamani Kirubakaran,
  • Chinnasamy Senthamarai Kannan,
  • Ranjith Raj Anbalagan,
  • Sivasubramanian Palanisamy,
  • Saleh A. Alfarraj,
  • Sulaiman Ali Alharbi,
  • Mohamed Abbas,
  • Shaeen Kalathil,
  • Mezigebu Belay

摘要

This research work utilizes machine learning, and taguchi method to optimize and predict the tensile performance of short carbon fiber reinforced polyamide (CF-PA) 3D printed by fused filament fabrication (FFF) technique. The process parameters play a major role in making novel material with help of FFF technology. In this research work, tensile samples are 3D printed based on different FFF process parameters, such as raster orientations, printer speed, and layer height as per ASTM D3039. While analysis of variance (ANOVA) confirms the statistical dominance of raster orientation (48.8% contribution), the framework’s novelty lies in leveraging a validated random forest model to uncover complex non-linear relationships and parameter interactions missed by traditional analysis. Methodologies including SHAP (Shapley Additive explanations) and partial dependence plots revealed a distinct non-linear effect of printer speed on tensile strength, peaking at 60 mm/s. Furthermore, the model was used to generate a process interaction map, demonstrating its utility as a predictive tool for exploring the design space beyond the initial experimental runs. The final SVR and random forest models demonstrated the highest predictive precision with the lowest average mean squared error under rigorous leave-one-out cross-validation (LOOCV). This work validates a cost-effective methodology for achieving deep process understanding from sparse datasets, accelerating the adoption of FFF-printed composites. The combination producing maximum strength determined through S/N analysis is stated as a printer speed at 60 mm/s with a 00 raster orientation having a layer height of 0.1 mm. A strong predictive tool can be developed out of a comparison among four machine learning models (ML), namely XGBoost, support vector regression (SVR), linear regression (LR), and random forest (RF) using leave-one-out cross-validation (LOOCV) as the primary validation strategy with mean squared error (MSE) being the major performance metric. The SVR and random forest models return the maximum predictive precision with the reduces average MSE. SEM morphology studies on samples support results obtained quantitatively by relating good mechanical properties to better fibre-matrix bonding and low porosity seen at optimal settings. This study is validating a low-cost method of producing CF-PA components, that ensure high performance with definite mechanical properties, expediting their use in highly demanding engineering applications.