Influence of artificial neural network architecture depth on surface roughness prediction in finish milling of 6061 aluminium alloy = Ảnh hưởng của độ sâu kiến trúc mạng nơ-ron nhân tạo đến dự đoán độ nhám bề mặt trong gia công phay tinh hợp kim nhôm 6061

Abstract

This paper investigates the influence of Artificial Neural Network (ANN) architecture depth on surface roughness prediction during finish milling of 6061 aluminium alloy. Experimental data were obtained from a Taguchi L27 design involving cutting speed, feed rate, and depth of cut. Twenty-one samples were used for training and six for internal testing, while six additional experiments outside the original design space were employed for external validation. ANN architectures containing one to four hidden layers were developed and evaluated using R², MAE, MAPE, and RMSE. All models achieved high accuracy on internal testing data, but the two-hidden-layer ANN provided the best performance with R² = 0.9970, MAE = 0.0124 µm, and RMSE = 0.0154 µm. External validation further confirmed its superior generalisation, achieving R² = 0.8715 and RMSE = 0.0333 µm. Deeper three- and four-layer architectures exhibited overfitting and reduced robustness. The results demonstrate that moderate network depth provides an effective balance between predictive accuracy, model complexity, and generalisation for machining applications with limited experimental datasets.