Abstract: Determining the scratch responses of metallic material plays a vital role in evaluating its resistance to scratching, friction and wearing. In the present study, finite element simulation was performed and the classical multi-layer perceptron (MLP) neural network method was used to explore the relationship between the input parameters in a scratch test (yield stress σ_y, strain hardening index n, interfacial friction coefficient μ_a, as well as the normal loading force F_n) and the scratch responses (apparent scratch depth d, scratch width w, and scratch tangential force F_t). The MLP neural network involved three parts: an input layer, hidden layers, and an output layer. Each layer in this model consists of many neurons, the neuronal behavior of which was described using the MP model (McCulloch-Pitts Model). Considering the time-consuming and costly process to obtain the dataset required for neural network training through experimental approach, finite element simulation was performed to obtain the scratch test dataset in this study. The plastic parameters of most metallic materials were covered by the selected range of values for yield stress and strain hardening index in the finite element simulation. A total of 960 scratch simulation cases are performed. The simulation dataset was randomly divided into three parts. Specifically, 60% of the data was included in the training set to learn the training neural network, 20% of the data, i.e, the validation set, was used to determine the optimal hyperparameters, and the remaining 20% in the dataset, i.e, the test set, was used to determine the generalization of the model. Herein, ReLU was chosen as the activation function because the linear correction unit (ReLU) activation function can speed up the training of the neural network model and effectively resolve gradient disappearance. The hyperbolic cosine loss function (Log-Cosh) was taken as the loss function. A search for 50,000 hyperparameter combinations was conducted on the open source platform pytorch using Optuna, and the models were assessed on the validation set, with the minimum Mean Absolute Percentage Error (MAPE) value as the optimal model hyperparameter. The best MLP fit for F_t, w and d predictions was 0.947, 0.986 and 0.97, respectively. Moreover, the MAPE of F_t falls below 11%, and that of w and d was smaller than 5.3% on both the training, validation, and test datasets, indicating the excellent predictability of the MLP. Scratch tests and tensile tests were conducted on 304 stainless steel, brass, and 18CrNiMo7-6 alloy steel to validate the trained MLP neural network. The results showed that the anti-scratch performance of the three materials used in the study range was ranked in the following order: 304 stainless steel > 18CrNiMo7-6 alloy steel > brass. This was because brass was less resistant to tensile deformation than 18CrNiMo7-6 alloy steel and 304 stainless steel. Although the yield stress of 304 stainless steel was lower than that of 18CrNiMo7-6 alloy steel, 304 stainless steel gradually outperforms 18CrNiMo7-6 alloy steel in the resistance to deformation with the increase in severity of tensile deformation. The relative errors of w and F_t of predicted scratch responsed from trained MLP neural network and scratch tests were within 10%, and the relative error of d was within 5%. The trained MLP neural network as proposed in this paper was applicable to make reasonable prediction of scratch responses for different metallic materials. Also, the results of this paper provided a viable solution to evaluating the anti-scratch performance of metallic materials.