Prediction of Aluminum Alloy Surface Roughness Through Nanosecond Pulse Laser Assisted by Continuous Laser Paint Removal

Abstract

Reducing surface roughness can enhance the mechanical properties of processed materials. The variation law of the aluminum alloy surface roughness induced by continuous-nanosecond combined laser (CL) with different continuous laser power densities and laser delay is investigated experimentally. A back propagation neural network (BPNN) coupled with a sparrow search algorithm (SSA) is employed to predict surface roughness. The nanosecond laser energy density, continuous laser power density and laser delay are input parameters, while the surface roughness is output parameter. The lowest surface roughness is achieved with completely paint film removed by the CL while the nanosecond laser energy density is 1.99 J/cm², the continuous laser power density is 2118 W/cm² and the laser delay is 1 ms. Compared to the original target and the target irradiated by nanosecond pulse laser (ns laser), the reductions in the surface roughness are 20.62% and 12.00%, respectively. The SSA-BPNN model demonstrates high prediction accuracy, with a correlation coefficient (R²) of 0.98628, root mean square error (RMSE) of 0.024, mean absolute error (MAE) of 0.020 and mean absolute percentage error (MAPE) of 1.30% on the test set. These results indicate that the SSA-BPNN demonstrates higher-precision surface roughness prediction with limited experimental data than BPNN. Furthermore, the findings confirm that the CL can effectively reduce surface roughness.

1. Introduction

Laser cleaning technology has been widely used in the processing and manufacturing industry due to its high controllability and significant efficiency [1,2,3]. This technology primarily relies on the thermal and mechanical effects of lasers to precisely remove surface contaminants on the surface of targets. By optimizing the laser parameters, it can effectively eliminate pollutants from various substrates, such as metals and plastics, without damaging the substrates. In addition, laser cleaning can achieve high precision surface treatment of different materials by accurately controlling the process parameters, which is especially suitable for the aviation industry and other fields with high requirements for surface treatment quality. Currently, laser cleaning technology has been successfully used in the field of rail transportation, which is used for paint removal of aluminum alloy car bodies and train wheels [4,5]. Numerous studies have further demonstrated that laser cleaning serves as a pretreatment method prior to welding, significantly enhancing welding quality [6,7,8]. The rough surface with sharp corners and notches can act as initiation points for crack formation under stress concentration. Consequently, reducing surface roughness helps minimize abrasion and prolongs the service life [9,10].

Recent research has increasingly focus on exploring the relationship between laser cleaning parameters and surface quality [11,12,13,14]. Recently, the research focus has gradually shifted to exploring the influence of laser cleaning parameters on the surface roughness of targets [15,16]. Guodong Zhu et al. [17] employed a 1064 nm Nd: YAG laser with pulse frequency of 7–15 kHz to clean 5A12 aluminum alloy, analyzing the effects of different laser powers and cleaning speeds on the surface roughness. The research results show that when the laser power is set to 98 W and the cleaning speed is 4.1 mm/s, the surface roughness reaches a minimum value of about 0.9 μm. Guangxing Zhang et al. [18] utilized a 1064 nm fiber laser with an average maximum output power of 100 W and a pulse frequency of 5 to 500 kHz to clean 5754 aluminum alloy, revealing a linear positive correlation between surface roughness and energy density. At the same energy density, the surface roughness will increase first and then decrease with the increase of the overlap ratio. Yunkai Li et al. [19] applied a high-repetition-frequency ns pulse laser with a wavelength of 1064 nm, pulse width of 10–400 ns, and an average power of 120 W to clean a 30 μm thick epoxy resin paint layer on the surface of 5083 aluminum alloy. It was found that the surface roughness fluctuated less after cleaning, and the corrosion resistance and coating adhesion were significantly improved. In studies of 1064 nm ns laser paint removal, Donghe Zhang et al. [20] studied the paint removal effect of ns laser with a pulse width of 100 ns on aluminum alloy. The results show that when the laser energy density is 6.4 J/cm², the surface roughness of the substrate is significantly reduced to 0.54 μm. Zejia Zhao et al. [21] studied the effect of ns laser with a pulse width of 140 ns on the paint removal of aluminum alloys. It was found that when the laser energy density was 1.66 J/cm², the surface roughness after laser cleaning was about 0.4 μm, which is close to the original aluminum alloy surface. It is indicated that the pulsed laser can achieve non-destructive paint removal from the substrate. Due to the thermal mechanism of continuous laser, it also has certain advantages in removing paint film from metal surfaces. A. Anthofer et al. [22] used a semiconductor continuous laser with a wavelength of 915–1030 nm, a power of 10 kW, and a spot area of 10 × 45 mm² to remove 350 μm thick epoxy resin paint film on the surface of concrete. Jason Provines et al. [23] used continuous laser to remove the coating on the surface of the steel bridge and found that in addition to the white coating, the removal rate of the coating by continuous laser ablation is 2–27 times faster than that of the pulsed laser ablation coating system. Based on the existing research results, it can be found that the research on surface roughness in the field of laser cleaning mainly focuses on a single laser cleaning mode, and most studies choose continuous laser or ns laser as the light source for laser cleaning. The main mechanisms of continuous laser and pulse laser are thermal effect and stress effect, respectively. By selecting suitable combination laser parameters, the paint film of aluminum alloys can be efficiently and completely removed. The surface roughness can also be reduced, simultaneously.

Extensive research has been conducted on surface roughness prediction in industrial processing [24,25,26,27]. However, focusing on the field of laser processing, the research on surface roughness prediction remains relatively limited. Li Wang et al. [28] employed a 1064 nm ns laser with a pulse width of 23.7 ns to irradiate alumina–copper ceramics and predicted the surface roughness after laser treatment using an artificial neural network model. The research results show that using Levenberg–Marquardt as the training function yielded optimal prediction accuracy. Chen Cao et al. [29] used the continuous laser to irradiate SiC ceramics and predicted the surface roughness by the support vector regression algorithm based on grey wolf optimization (GWO-SVR). By comparing the performance of the model before and after optimization, it was found that the relative error of the optimized model was reduced by 1.91%, which fully demonstrates that the GWO-SVR can accurately predict the surface roughness of SiC ceramics during laser processing. Nevertheless, the research on surface roughness prediction in the field of laser cleaning has not been reported.

The decrement of surface roughness can improve the mechanical properties of the target and significantly improve the welding performance. With the aim of reducing the aluminum alloy surface roughness induced by CL cleaning epoxy resin paint film, the variation law of the aluminum alloy surface roughness with different continuous laser power densities and laser delay is studied. The effect of CL on surface roughness compared with ns laser is clarified. Moreover, the BPNN and the SSA-BPNN are developed to predict the surface roughness. The nanosecond laser energy density, continuous laser power density and laser delay are input parameters, while the surface roughness is the output parameter. This provides theoretical guidance for improving the target surface properties in laser welding processing. This study provides a quick and accurate method for predicting surface roughness.

2. Materials and Methods

2.1. Experiment

The experimental setup for the CL paint removal on aluminum alloy is shown in Figure 1. The CL paint removal experimental system includes a continuous laser (IPG Photonics, Marlborough, MA, USA), an ns laser (Beamtech, White Rock, BC, Canada), two photodetectors (THORLABS, Newton, NJ, USA), a DG645 pulse delay generator (Stanford Research Systems, Sunnyvale, CA, USA), and an oscilloscope (Tektronix, Beaverton, OR, USA). The continuous laser has a wavelength of 1070 nm and a duration of 1 s and a spot diameter of 2.5 mm. The ns pulse laser has a wavelength of 1064 nm and a pulse width of 12 ns and a spot diameter of 1.5 mm. The ns laser beam and the continuous laser beam are focused on the target at the same point using two lenses (L₁ and L₂) with a focal length of 200 mm. The spatial angle between the two laser beams is 10 degrees. After the continuous laser is emitted, photodetector 1 detects the scattered light from the continuous laser and transmits the rising edge signal to the DG645 pulse delay generator. The DG645 controls the delay time of the CL and triggers the ns laser. Photodetector 2 transmits the detected waveform of the CL to the oscilloscope for real-time monitoring of the delay time. The delay time between the ns laser and the start point of the continuous laser is defined as Δt. The ns laser energy density is defined as E_ns. The continuous laser power density is defined as P_C. The ns laser energy density-continuous laser power density is defined as the CL energy parameters E_CL. The target is an aluminum alloy substrate (6061 model) with a thickness of 1 mm, the coating is an epoxy resin layer with a thickness of 26 μm, and the main component of the epoxy resin coating is bisphenol A epoxy resin. Bisphenol A epoxy resin is a high molecular compound which is prepared by condensation of bisphenol A and epichlorohydrin under alkaline conditions, washing and desolvation. The experimental sample size is 40 mm × 40 mm (length × width).

In this study, a three-dimensional surface automatic measuring instrument (IFM G3, Bruker Alicona, Graz, Austria) with 20 times magnification is used to measure the surface roughness and surface morphology. The surface roughness is calculated by the average of the six line roughness. The six line roughness are three horizontal and three vertical lines of roughness in the laser removal area, respectively, which are distributed as shown in Figure 1. Each group of experiments was repeated three times.

2.2. BPNN and SSA

2.2.1. BPNN

BPNN is a multilayer feedforward neural network consisting of an input layer, an output layer and one or more hidden layers. The data are imported into the input layer, weighted by the hidden layer neurons, and finally transmitted to the output layer for result. If the output deviation exceeds the predefined threshold, the network will activate the error back-propagation mechanism. In this process, the error starts from the output layer and propagates forward layer by layer. Each layer of neurons dynamically adjusts the connection weights and thresholds in the direction where the error function declines fastest. The training process will stop until the output deviation reaches the preset condition, or when the maximum number of iterations is reached. The specific calculation rules of the BPNN are determined by Equations (1) and (2).

$$h=f\left({\displaystyle \sum }_{i=1}^4{\omega }_{2}f\left({\displaystyle \sum }_{i=1}^4{\omega }_1{x}_i+b_1\right)+b_2\right)$$

(1)

$$y=g\left({\displaystyle \sum }_{j=1}^5{\omega }_{3}y+b_3\right)$$

(2)

where x_i is the input, y is the output, h is the hidden layer neuron, b_j and ω_j are the connection thresholds and weights between the input layer and the hidden layer, the first hidden layer and the second hidden layer, and the hidden layer and the output layer, respectively. f( ) as the hidden layer transfer function is set to the tansig function; g( ) as the output layer transfer function is set to the purelin function.

The structure of the BPNN model is shown in Figure 2. The input layer is set to consist of three neurons, which are E_ns, P_C and Δt. The hidden layer is set into two layers, each layer consisting of 5 neurons. The output layer consists of 1 neuron, and its output result is surface roughness. During the training process, the number of training is 10,000, the learning rate is 0.05, and the minimum error of the training target is 0.00001. The 1st to 19th data are set as the training set, and the 20th to 25th data are set as the test set as shown in

Table 1. Sample set.

Sample Number	Ens/J/cm2	PC/W/cm2	Δt/ms
1	0	0	0
2	1.99	0	0
3	0	3621	0
4	0	6564	0
5	0	8035	0
6	0	9733	0
7	0	10,865	0
8	1.99	1711	0
9	1.99	1711	1
10	1.99	1711	10
11	1.99	1711	100
12	1.99	1711	500
13	1.99	1711	1000
14	1.99	1992	0
15	1.99	1992	1
16	1.99	1992	10
17	1.99	1992	100
18	1.99	1992	500
19	1.99	1992	1000
20	1.99	2118	0
21	1.99	2118	1
22	1.99	2118	10
23	1.99	2118	100
24	1.99	2118	500
25	1.99	2118	1000

. The surface roughness prediction model of aluminum alloy in this study is only for epoxy resin paint film with 26 μm thickness.

2.2.2. SSA-BPNN

Although BPNN exhibits powerful parallel computing and nonlinear modeling capabilities, it also has certain limitations when dealing with practical problems. Specifically, its generalization ability is relatively weak, often requiring extensive training datasets to achieve optimal prediction performance. Additionally, BPNN is extremely prone to getting stuck in local optima [30]. In view of this, we employ the SSA to optimize the BPNN, enhancing its robustness and predictive accuracy. The optimized model is applied to predict surface roughness after laser cleaning, and the specific operation process is shown in Figure 3.

The SSA was inspired by the foraging behavior and anti-predation behavior of sparrows. In the natural environment, the sparrow population shows obvious division of labor characteristics in foraging and anti-predation activities, mainly including three roles: discoverer, follower and alerter. The SSA makes group decisions based on the position information of these three roles and the target objective, and searches for the optimal solution by constantly updating their positions.

The location update description of the discoverer is shown in Equation (3):

$$X_{i,j}^{t+1}=\left\{\begin{array}{c}{X}_{i,j}^t·exp\left(-{\displaystyle \frac{i}{\alpha ·ite{r}_{max}}}\right) if R_2<ST\\ X_{i,j}^t+Q·L if R_2\ge ST\end{array}\right.$$

(3)

where $t$ represents the number of current iterations; $X_{i,j}^t$ represents the position information of the $i$th sparrow in the $\mathrm{j}$ dimension at $t$ time; $\alpha ,Q$ is a random number; $ite{r}_{max}$ is a constant, represents the maximum number of iterations; $L$ denotes a matrix of $1\times d$, where each element in the matrix is 1. When $R_2\ge ST$, indicating that there may be danger, the discoverer will lead the other sparrows to the safe area to continue foraging.

The location update description of the follower is shown in Equation (4):

$$X_{i,j}^{t+1}=\left\{\begin{array}{c}Q·exp\left(-{\displaystyle \frac{X_{worst}^t-X_{i,j}^t}{i^2}}\right) if i>{\displaystyle \frac{n}{2}}\\ X_p^{t+1}+\left|X_{i,j}^t-X_p^{t+1}\right|·A^{+}·L otherwise\hfill \end{array}\right.$$

(4)

Among them, $X_{worst}^t$ is the position with the worst fitness in the population at $t$ time; $X_p^{t+1}$ is the optimal position of fitness among the discoverers at $t+1$ time; $A$ represents a $1\times d$ matrix. When $i>{\displaystyle \frac{n}{2}}$, it means that the $i$th follower has the worst fitness value and foraging is not completed; it will fly elsewhere to forage. Conversely, it indicates that it forages around sparrows with higher fitness.

The position update description of the alerter is shown in Equation (5):

$$X_{i,j}^{t+1}=\left\{\begin{array}{c}{X}_{best}^t+\beta \left|X_{i,j}^t-X_{best}^t\right| if f_i>f_g\\ X_{i,j}^t+K·\left({\displaystyle \frac{\left|X_{i,j}^t-X_{worst}^t\right|}{\left(f_i-f_w\right)+\epsilon }}\right) if f_i=f_g\end{array}\right.$$

(5)

where $X_{best}$ is the optimal position in the population at $t$ time; $f_i$ is the fitness value representing the $i$th sparrow. $f_g$ and $f_w$ are the current global optimal and worst fitness values, respectively. $\epsilon$ is a very small constant to avoid a zero denominator. When $f_i=f_g$, the alerter sparrows will approach other sparrows to prevent predation.

The SSA has a strong global optimization capability, which can optimize the initial weights and biases of the BPNN to improve its efficiency and accuracy. The initial population size of the SSA is 30; the maximum evolutionary generation is 50; the safety value ST is 0.6; the proportion of discoverer PD was 0.2; the proportion of alerter sparrows SD is 0.1.

2.2.3. Model Evaluation Indicators

To comprehensively evaluate the prediction model, MAE, MAPE, RMSE and R² are selected as evaluation indicators. The calculation equations of these indicators are shown in Equations (6)–(9). The numerical value of the three evaluation indexes of MAE, MAPE and RMSE is negatively correlated with the accuracy of the model, that is, the smaller the value, the smaller the error generated by the model in the prediction process. For R², its value range is between 0 and 1. When the value of R² approaches 1, it shows that the model has a good fitting effect on the observed data and can accurately reflect the relationship between input and output, which proves that the prediction model has high accuracy.

$$MAE={\displaystyle \frac{1}{N}}{{\displaystyle \sum }}_{i=1}^N\left|y_p-y_t\right|,$$

(6)

$$MAPE={\displaystyle \frac{1}{N}}{{\displaystyle \sum }}_{i=1}^N\left|{\displaystyle \frac{y_p-y_t}{y_t}}\right|,$$

(7)

$$RMSE=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum }_{i=1}^N{(y_p-y_t)}^2},$$

(8)

$$R^2={\displaystyle \frac{\left[{{\displaystyle \sum }}_{i=1}^N{\left(y_t-\overline{y}\right)}^2\right]-\left[{{\displaystyle \sum }}_{i=1}^N{\left(y_t-y_p\right)}^2\right]}{{{\displaystyle \sum }}_{i=1}^N{\left(y_t-\overline{y}\right)}^2}},$$

(9)

In the equation, N is the number of samples, $y_p$ is the predicted value, $y_t$ is the actual measured value, and $\overline{y}$ is the average value of the sample.

3. Results and Discussion

3.1. Experimental Surface Roughness

3.1.1. The Surface Morphology of Target Induced by Laser Paint Removal

The comparative morphology of the target before and after laser paint removal is shown in Figure 4. As demonstrated in Figure 4a–d, the paint film can be completely removed by the CL parameters of E_CL = 1.99 J/cm² − 2118 W/cm² and Δt = 0 and 1 ms. While Δt = 10 ms, near-complete removal is observed, though with residual paint film adhering to the target, which is shown in Figure 4e. As shown in Figure 4f–h, the paint film is seriously attached to the surface of the aluminum alloy while Δt ≥ 100 ms. This phenomenon can be attributed to the ablation mechanism of continuous laser allowing the paint film to melt and adhere to the target surface. Under these conditions, the stress and plasma impact force induced by the ns laser cannot remove the paint film completely.

3.1.2. The Surface Roughness of the Target Induced by Laser Paint Removal

Figure 5 illustrates the variation in the surface roughness with Δt. When E_CL = 1.99 J/cm² − 1711 W/cm², the surface roughness of the target decreases with the increase of Δt. When E_CL = 1.99 J/cm² − 1992 W/cm² and E_CL = 1.99 J/cm² − 2118 W/cm², both trends are consistent. The surface roughness decreases as the Δt increases while 0 ≥ Δt ≥ 10 ms. The surface roughness reaches the lowest and the paint film is almost removed while Δt = 10 ms. Combined with Figure 4e, this optimal condition corresponds with near-complete paint removal, which we attribute to laser-induced plasma effects, the combined action of surface ionization and thermal stress coupling promotes paint film delamination from the aluminum alloy [31], while residual paint particles fill in the gaps on the metal surface, resulting in a smooth surface. The surface roughness increases while Δt = 100 ms. This may be due to the melting phenomenon of the paint film induced by the irradiation of continuous laser. The shocking effect of the ns laser will cause the molten paint film to splash, and the accumulation of splash will increase the surface roughness.

As shown in Figure 4c–e, when E_CL = 1.99 J/cm² − 2118 W/cm², the aluminum alloy is almost exposed to the surface irradiated by CL while Δt = 0, 1 and 10 ms. The increment of the surface roughness is shown in Figure 6. The original surface roughness of the target is 1.94 μm, and the surface roughness induced by ns laser is 1.75 μm. Compared with the original target and the target irradiated by ns laser, the surface roughness induced by the CL is greatly reduced while Δt = 10 ms. However, there is still a slight paint film on the target surface as shown in Figure 4e. When Δt = 1 ms, the paint film is completely removed. The surface roughness induced by CL is 12.00% reduction in roughness compared to ns laser, and a 20.62% reduction in roughness compared to the original target.

3.2. ANN Prediction for Surface Roughness

The surface roughness prediction models based on BPNN and SSA-BPNN are shown in Figure 7. It reveals that the SSA-BPNN predicts the surface roughness more closely to the actual values than BPNN in both the test and training process.

The Coefficient of Determination for both prediction models are shown in Figure 8. For BPNN, The R² of the training set (0.97354) demonstrates relatively higher than the R² of the test set (0.90538) as shown in Figure 8a. It shows that BPNN indicates limited generalization capability and low prediction accuracy on unknown test data. In contrast, the SSA-BPNN achieves R² of the training set (0.99851) and the R² of the test set (0.98628), demonstrating excellent agreement between predicted and actual values across both datasets, which is shown in Figure 8b. These results confirm that the SSA-BPNN maintains consistently high predictive performance while effectively overcoming the generalization limitations observed in the BPNN.

The evaluation indexes of BPNN and SSA-BPNN are shown in

Table 2. Evaluation indexes of prediction model.

Model	RMSE	MAE	MAPE
BP	0.033	0.05	6.45%
SSA-BP	0.024	0.020	1.30%

. The RSME, MAE and MAPE values of BPNN are higher than SSA-BPNN. It can be observed that there are extreme errors in the BPNN predictions due to the RMSE for BPNN being greater than SSA-BPNN; therefore, its error will be relatively high. The gap between the predicted and actual values of SSA-BPNN is relatively smaller. And the SSA-BPNN has higher prediction accuracy than BPNN. Therefore, the SSA-BPNN has more reliable surface roughness predictions compared to BPNN.

Figure 9a indicates the comparison of the surface roughness between actual values, BPNN predictions and SSA-BPNN predictions; the SSA-BPNN predictions exhibit significantly better agreement with the actual values than BPNN. The relative errors of surface roughness predictions are shown in Figure 9b; the relative errors of SSA-BPNN are basically lower than BPNN. The surface roughness predictions, relative errors and absolute errors of BPNN and SSA-BPNN are shown in

Table 3. Surface roughness predictions, relative errors and absolute errors.

Sample Number	Actual Value/μm	Predicted Value/μm		Relative Error/%		Absolute Error/μm
		BP	SSA-BP	BP	SSA-BP	BP	SSA-BP
20	1.78	1.70365	1.74226	4.289	2.12	−0.07635	−0.03774
21	1.54	1.6656	1.58031	8.156	2.617	0.1256	0.04031
22	1.28	1.32904	1.27634	3.831	0.286	0.04904	−0.00366
23	1.65	1.65823	1.64211	0.499	0.478	0.00823	−0.00789
24	1.33	1.31196	1.34685	1.357	1.266	−0.01804	0.01685
25	1.22	1.19728	1.20726	1.862	1.044	−0.02272	−0.01274

. As shown in Figure 9a, for samples 23–25, the predictions of both models are basically within the error bar range. However, for samples 20–22, only the SSA-BPNN maintains predictions within acceptable error margins, while BPNN predictions exceed this range. The absolute errors of BPNN are −0.07635, 0.1256 and 0.04904, the absolute errors of SSA-BPNN are −0.03774, 0.04031 and −0.00366, and the absolute errors of SSA-BPNN are lower than BPNN. It reveals that after the BPNN is optimized by SSA, the error is greatly reduced, which solves the problem that it falls into the minimum and affects the prediction accuracy.

4. Conclusions

In this paper, the variation of surface roughness with the E_CL and △t of CL is studied, and the surface roughness prediction model is established based on BPNN and SSA-BPNN. The Ens, P_C and △t are input parameters, while the surface roughness is the output parameter. The following are the primary conclusions:

(1)

Compared to the single ns laser, the CL can reduce aluminum alloy surface roughness and completely remove the paint film while ECL = 1.99 J/cm² − 2118 W/cm², △t = 1 ms. The surface roughness reduction is from 1.75 µm to 1.54 µm, a decrement of 12.00% and a 20.62% roughness reduction compared to the original target.

(2)

The surface roughness predictions of SSA-BPNN fit better with the test set than BPNN. The R², RMSE, MAE and MAP are 0.98628, 0.024, 0.020 and 1.30%, respectively. It is indicated that the SSA-BPNN has higher prediction accuracy and a higher reliability for surface roughness. This provides a theoretical reference for both surface quality enhancement and laser parameter optimization in industrial applications.