**4. Discussion**

The results obtained with Test 2 allowed for concluding something already expected, namely, that by increasing the number of data provided to the ANN, the learning process leads to high accuracy results, and thus the amount of data introduced in the future should be even higher, allowing for more accurate predictions. This directly results from the observations of Table 4 and Figure 5. This trend was previously argued by Tay and Cao [56], namely, that ANN results depend to a large extent on the amount of data provided to the system, being a time-consuming task.

Based on the results previously described, it can be stated that many network architectures have provided very accurate results, with eight architectures showing errors lower than 3.15% (96.85% accuracy), which can be considered an excellent result in terms of machining cost prediction. This result is less than those previously obtained by other researchers [57], who reported average accuracies of about 98.5% using ANNs, which has been shown to be a more accurate method to detect the different blends of fuel than RSM (response surface methodology) methods. Moreover, another study [58] about predictions of AISI 1050 steel machining performance has shown that ANNs presented the most accurate prediction value (92.1%), which is below the accuracy provided by the eight different architectures used in this work. In that work [58], it remains clear that ANN can provide more accurate values than other prediction techniques, such as the adaptive neuro-fuzzy inference system (ANFIS), which provided an accuracy of 73% compared to the 92.1% achieved using ANN. Saric et al. [55] also claimed an error of 2.03% using back-propagation neural networks in the estimation of CNC machining times. However, using self-organizing map neural networks, the results are less accurate, showing errors of about 10.05%. Ning et al. [52] used as a training dataset 21,943 features, and 4338 for validation, obtaining an accuracy of 97.7%, which is very similar to that achieved through this work. Thus, it is expected that a higher number of features used for training of the model can contribute to a better accuracy of the predictions, but too much data cannot contribute to accuracy improvement and can cause higher computing time, which can be unnecessary.

Regarding Table 5, the best results were obtaining using as input variables the volume of removed materials (V) and the total number of elements of each feature (Q). Indeed, the second most accurate result was obtained using the same variables plus the machined

surface area (A). However, adding this variable, the results started to degrade. Among the other combinations considered in this work, only the Q + A combination presented acceptable results, together with the combination Q + At. Thus, it is possible to infer that the volume of removed materials plays an important role in terms of keeping the accuracy as high as possible, which can be combined with V for the best results and combined with A for close to the best results.
