What question did this study set out to answer?

The aim is to optimize neural networks for predicting depth of cut in abrasive water jet machining of Ti6Al4V material.

July 3, 2026Open Access

Modelling and computational optimization of different neural network architectures for prediction of depth of cut in abrasive water jet machining of Ti6Al4V

Key Points

The aim is to optimize neural networks for predicting depth of cut in abrasive water jet machining of Ti6Al4V material.
Used Taguchi-based L27 Orthogonal array for experimentation with five parameters: water pressure, traverse speed, nozzle diameter, abrasive mass flow rate, and orifice size.
Developed two neural network architectures: single hidden layer and deep neural network, with variations in neurons and layers.
Employed four activation functions and evaluated models using RMSE and R2 to determine accuracy.
NN D outperformed with a 12.5% reduction in testing RMSE and 4.2% improvement in R2 compared to NN 1.
Gaussian activation function was the most effective, reflecting the jet energy distribution accurately.
SEM analysis indicated distinct machining behaviors for varying depths of cut, confirming predictions about water pressure and traverse speed influence.

Abstract

The current work presents an optimized neural network for predicting depth of cut (DoC) in the abrasive water jet machining (AWJM) of Titanium Ti6Al4V Grade 5 material. Five process parameters were used for experimentation i. e. , water pressure (W p), traverse speed (T s), nozzle to orifice diameter (N/O dia), abrasive mass flow rate (A mf), and abrasive orifice size (A os). Experiments were carried out using the Taguchi based L 27 Orthogonal array, resulting in the development of a regression equation describing the process behaviour. The DoC was modelled using two neural network architectures: a single hidden layer network (NN 1), where neurons were varied from 1 to 10, and a deep neural network (NN D), where both neurons and hidden layers were varied to determine the optimum network configuration. Four different activation functions, namely, Sigmoidal, Gaussian, Tanh, and Linear functions were employed to perform the optimization. The dataset for training and testing the neural network models was generated using the regression equation based on experimental data, rather than direct experimental measurements. The optimization of both the neural network architectures was evaluated using two performance metrics, i. e. , root mean squared error (RMSE) and co-efficient of determination (\ (\: R^2\) ). The results concluded that both the neural network (NN) models were able to accurately predict DoC, however, the NN D was found to be better in terms of accuracy with a 12. 5% reduction in testing RMSE and 4. 2% improvement in testing (\ (\: R^2\) ). Amongst the activation functions used, the Gaussian functions outperformed the other functions across both neural network architectures, as its symmetric response reflects the jet energy distribution peaking at the centre and decaying radially, thereby effectively capturing the non-linear penetration behavior. SEM analysis supported the computational results, showing micro-cutting with stable grooves for machined surfaces with higher DoC, whereas lower DoC experienced micro-ploughing and ridge formation. Contour plots further revealed that DoC in AWJM of Ti6Al4V is primarily governed by water pressure and traverse speed, thereby confirming the capability of the optimised NN D architecture to capture the underlying physical mechanisms of AWJM.

Ask AI

Helpful

Bookmark

View Full Paper