Hyper parameter settings for ML Algorithms.

3D printing has brought significant changes to manufacturing sectors, making it possible to produce intricate, multi-layered designs with greater ease. This study focuses on optimizing the compressive strength (CS) of functionally graded multi-material (PLA/Almond Shell Reinforced PLA) which is fabricated with the aid of the FFF process, a widely used additive manufacturing technique. Six different machine learning models (ML) were utilized to estimate CS using key process parameters, namely print speed (PS), layer height (LH), and printing temperature (PT). Among six ML models, Polynomial Regression (PR) performed best, with an R2 of 0.88 and the lowest error metrics (MAE = 1.38, RMSE = 1.9, MSE = 3.6). SHAP analysis indicated that PS is the most influential parameter, followed by LH. PR predicted optimal parameters (PS = 19 mm/s, LH = 0.1 mm, PT = 216°C) and yielded a predicted CS of 36 MPa, which was experimentally validated as 34.8 MPa with a low error of 3.44%. Also, the PR outperformed the traditional Taguchi method, which predicted a CS of 33.74 MPa, showing a 7.5% improvement and lower error. This demonstrates that PR-based ML optimization offers better accuracy and improved mechanical performance, making these FGMs suitable for various consumer applications.

3D打印（3D printing）已为制造业带来深刻变革，使复杂多层结构的高效制备成为可能。本研究聚焦于优化功能梯度多材料（functionally graded multi-material, FGM）的抗压强度（compressive strength, CS），该材料采用当前广泛应用的增材制造技术——熔融沉积成型（Fused Filament Fabrication, FFF）制备，具体为聚乳酸/杏仁壳增强聚乳酸（PLA/Almond Shell Reinforced PLA）体系。研究选取打印速度（print speed, PS）、层厚（layer height, LH）与打印温度（printing temperature, PT）三个核心工艺参数，采用6种机器学习（machine learning, ML）模型对抗压强度进行预测。在6种模型中，多项式回归（Polynomial Regression, PR）表现最优，其决定系数（R²）达0.88，且误差指标最低（平均绝对误差MAE=1.38，均方根误差RMSE=1.9，均方误差MSE=3.6）。SHAP分析（SHapley Additive exPlanations）结果显示，打印速度为影响抗压强度的最关键参数，其次为层厚。多项式回归模型预测得到最优工艺参数组合：打印速度19 mm/s、层厚0.1 mm、打印温度216°C，对应的预测抗压强度为36 MPa；经实验验证，实际抗压强度为34.8 MPa，误差仅为3.44%。此外，多项式回归模型的表现优于传统田口方法（Taguchi method），后者预测的抗压强度为33.74 MPa，相较之下提升了7.5%且误差更低。本研究表明，基于多项式回归的机器学习优化方案具备更高的预测精度与更优的力学性能，使得这类功能梯度多材料可适配多种民用应用场景。