Optimization of Aggregate Characteristic Parameters for Asphalt Binder—Aggregate System under Moisture Susceptibility Condition Based on Random Forest Analysis Model

Abstract

Water damage to asphalt pavements is a common occurrence that lowers the quality of service they can offer and causes several traffic problems. The loss of adhesion characteristics in the system of the asphalt binder and aggregate is the primary source of the problem of water damage in asphalt mixes. A number of things, including the impact of aggregate characteristics on asphalt binder—aggregate systems’ adhesion characteristics, have been proven. Through the use of random forest analysis, this study seeks to maximize the screening of aggregate characteristic factors. In this research, the morphology characterization, chemical composition, and phase composition of the five aggregates were first studied, and their relevant characteristic parameters were calculated. A method of engineering evaluation of the resistance of asphalt mixtures to water damage was used to assess the water susceptibility of the five asphalt binder—aggregate systems. Next, utilizing the fuzzy comprehensive evaluation analysis approach, a thorough study of the water susceptibility of the five asphalt binder—aggregate systems was conducted. Finally, sensitivity analysis of the aggregate characteristic parameters was carried out by a random forest analysis model, so as to achieve the optimal screening of the aggregate characteristic parameters. The results showed that, during sensitivity analysis of each parameter of aggregate properties using random forest analysis, the SiO₂ content of the aggregate had the highest importance, and the roughness had the highest importance among the morphology characterizations. The water susceptibility of the asphalt binder—aggregate system could be expressed by the SiO₂ content and roughness of the aggregate characteristic parameters.

1. Introduction

In road development and construction, asphalt pavement is widely used because of its excellent performance in anti-slip and in noise reduction [1,2]. Aggregate, asphalt binder and voids make up the asphalt mixture, where the asphalt binder holds the aggregate particles together [3,4]. A large number of pavement distresses can exist during the service life of asphalt pavements, including peeling, slurry, potholes, and cracks. Asphalt pavements may become less stiff and structurally unsound due to these pavement distresses, which may affect how long they last [5,6]. Water damage in asphalt mixtures is mostly caused by a decline in the asphalt binder and the aggregate system’s bonding capacity [7,8].

Since aggregates and asphalt binder make up the bulk of asphalt mixtures, it is essential to start with these two elements of the asphalt binder—aggregate system when assessing the water susceptibility of asphalt mixtures [9,10]. In order to determine the surface free energy parameters of five aggregates and two asphalt binders, Bhasin et al. used the surface free energy theory. They also confirmed that there is a connection between the asphalt binder’s various components and the mineral element composition of the aggregates, as well as the surface free energy parameters of the asphalt binder—aggregate system [11]. Exploring the different components of asphalt binder, Curtis et al. found that the degree of influence on the adhesion of asphalt binder—aggregate systems was in descending order of carboxyl, sulfoxide, pyrimidine, phenol, pyrrolidine, and ketone [12]. Exploring the different components of asphalt binder, Chen found that resin and asphaltene in asphalt contributed more to the adhesion properties than aromatic and saturated fractions. From the asphalt properties, Zhou et al. found that the acid value of asphalt binder was the most influential factor in the adhesion performance of asphalt mixes [13]. Feng et al. found that there was no good correlation between asphalt binder viscosity and adhesion performance of an asphalt—aggregate system by the Brookfield rotational viscosity test. They found the ratio of resin to asphaltene could evaluate the adhesion performance of asphalt mixtures [14].

The adhesion properties of the asphalt binder—aggregate system are largely determined by the aggregate properties since aggregate makes up about 95% of the weight of the entire asphalt mixture [15]. In natural conditions, an asphalt mixture under the action of water, the asphalt film on the surface of the aggregate will spontaneously peel off, which also indicates that the aggregate has hydrophilic properties. Because aggregate adhesion to water is often stronger than aggregate adhesion to asphalt binder, the importance of aggregate properties for the research of asphalt binder—aggregate system adhesion qualities cannot be overstated [16]. Walsh et al. determined that the kind of aggregate had a substantial influence on the adhesion after using the net adsorption method to test the adhesion between different aggregate types and asphalt binder [17]. Many academics have conducted in-depth studies on the aggregate properties because an experimental study in NCAT demonstrated that the influence of aggregate properties on the adhesion properties of the asphalt binder—aggregate system is significantly greater than that of asphalt binder in asphalt mixtures [18]. According to Elphingstone et al., the SFE characteristics of the asphalt binder—aggregate system were computed to forecast how resistant asphalt mixes would be to water damage. Additionally, it was discovered that the aggregates’ impact on the adhesive qualities of the asphalt binder—aggregate combination was noticeably stronger than that of the asphalt binder [19]. Bagampadde selected 11 types of aggregates in asphalt mixtures as independent variables. The mineral phase composition and chemical elemental composition of different types of aggregates were measured to evaluate the water damage resistance of asphalt mixes. It was found that alkali metal elements, such as K and Na, in the aggregates were more detrimental to the water damage resistance of asphalt mixes compared to Mg and Fe elements. The difference in water damage resistance in asphalt mixtures was obtained mainly from the difference in aggregate types [20]. The angularity, sphericity, and surface roughness of aggregates were studied by Guo et al. using an aggregate image acquisition system. They discovered that the water sensitivity of asphalt mixtures was significantly related to aggregate morphology [21]. Valdés et al. found that the geometric properties of aggregates had a binder strength to some extent [22].

In conclusion, an increasing number of scholars are gradually shifting their research focus to the aggregates in asphalt binder—aggregate systems. By examining the aggregate properties, it is possible to determine how the aggregate properties affected the adhesion characteristics of asphalt binder—aggregate systems. The current research’s findings also supported the hypothesis that aggregate characteristics affect how well asphalt binders and aggregates adhere to one another. However, most scholars, when studying the influence of aggregate properties on the adhesion performance of asphalt binder—aggregate systems, only qualitatively evaluate the influence of aggregate properties on the adhesion performance of the asphalt binder—aggregate system, lacking the correlation analysis between different aggregate property parameters, and cannot determine the importance ranking of aggregate property parameters. Therefore, in order to study in depth the connection between different aggregate characteristic parameters, an appropriate importance analysis method must be used to optimize the screening of different aggregate characteristic parameters.

In this research, the morphology characterization, chemical composition, and phase composition of the five aggregates were first studied. The five asphalt binder—aggregate systems’ water susceptibility was assessed using the engineering approach for assessing the resistance of asphalt mixtures to water damage. Their pertinent characteristic parameters were determined. Then the water susceptibility of the five asphalt binder—aggregate systems was evaluated comprehensively by using the fuzzy comprehensive evaluation method. Finally, sensitivity analysis of the aggregate characteristic parameters was carried out by a random forest analysis model, so as to achieve the optimal screening of the aggregate characteristic parameters.

2. Raw Materials and Testing Methods

2.1. Experimental Design

Figure 1 depicts the experiment design flowchart for this investigation, together with the test procedures and raw materials used.

2.2. Raw Materials

(1)

Asphalt Binder

For this experiment, the #90 virgin asphalt binder (graded according to penetration values) was employed to avoid having modified asphalt binders affect the ability of asphalt binder—aggregate systems to adhere [23]. An overview of the properties of the unprocessed #90 asphalt binder can be found in

Table 1. Properties of asphalt binder.

Properties		Value	Test Methods
Penetration (25 °C, 100 g, 5 s) (0.1 mm)		88.7	T0604
Softening point (°C)		46.8	T0606
Ductility (5 cm/min, 15 °C) (cm)		>100	T0605
Brookfield viscosity (135 °C) (mPa·s)		≤3	T0619
After RTFOT (163 °C, 85 min)	Quality change (%)	±1	T0608
	Residual penetration ratio (25 °C) (%)	≥65	T0604
	Residual ductility (25 °C) (cm)	≥8.0	T0605

.

(2)

Aggregate

As aggregate forms up to 95% of the weight of the asphalt mixture, the surface characteristics of the aggregate have a significant impact on how well the aggregate and asphalt cling to one another. For this analysis, type A limestone, type B limestone, basalt, granite, and diabase were the five types of aggregates taken into account.

Table 2. Properties of aggregates.

Test	Type A Limestone	Type B Limestone	Basalt	Granite	Diabase	Test Methods
Apparent density (g/cm3)	2.792	2.815	3.002	2.842	2.983	T0330-2005
Hygroscopic rate (%)	0.5	0.34	0.6	0.62	0.51	T0305-1994
Crush value (%)	15.9	12.7	11.3	10.1	12.4	T0316-2005
Los Angeles abrasion (%)	14.3	13.6	12.4	13.5	6.8	T0317-2005
Polishing value	43	44	42	44	45	T0321-2005

provides a list of these five aggregates’ key properties. The five aggregates’ morphology was evaluated using scanning electron microscopy (SEM) analysis. The Field Emission Scanning Electron Microscope FE-SEM is manufactured by Carl Zeiss, Germany. X-ray diffraction (XRD) and X-ray fluorescence (XRF) were employed to analyze the phase compositions and chemical compositions.

2.3. Testing Methods

2.3.1. Modified Boiling Water Test

The adhesion quality of the asphalt binder—aggregate system was evaluated using the boiling water test. In the beginning, five aggregates with a particle size of 13.2 mm were dried for an hour at 105 ± 5 °C in an oven and the mass of the aggregates (m₁) was weighed. Then, for 45 s, these aggregates were immersed in a liquid asphalt binder. The coated particles were then dried off by hanging them at a temperature of 25 °C for 15 min and the mass of the coated particles (m₂) was weighed. Following that, these particles were cooked in water for three minutes and their mass (m₃) was weighed. The test results were then visually categorized into five grades based on the degree of peeling. The indicator of the peeling rate W_b (%) was chosen to better measure this adhesion loss. Equation (1) was used to get the peeling rate W_b (%).

$$W_b=\frac{m_3-m_2}{m_2-m_1}\times 100$$

(1)

2.3.2. Immersion Marshall Test

The immersion Marshall test was used to determine how susceptible the compacted asphalt mixture was to moisture. The type AC-13 aggregate gradation was chosen. Two groups of typical Marshall specimens with dimensions of 101.6 mm by 63.5 mm were produced using the Marshall compaction molding technique. The first group was tested using the conventional test procedure after being submerged in a 60 °C water bath for 30 min to ascertain its susceptibility (MS). The second group was submerged in a 60 °C water bath for 48 h. The second group’s susceptibility (MS₁) was then evaluated. The residual susceptibility (MS₀) under water immersion was calculated using Equation (2).

$$M{S}_0=\frac{M{S}_1}{MS}\times 100$$

(2)

2.3.3. Freeze—Thaw Splitting Test

In this experiment, the freeze—thaw splitting test was used to look into the water susceptibility of the compacted asphalt mixture. Gradation type AC-13 was applied to the asphalt mixture. For this test, two sets of standard Marshall specimens were gathered. The second group was kept at −18 °C for 16 h before spending another 24 h in a water bath at 60 °C. One member of the groups was kept at ambient temperature (25 °C). All specimens’ maximum load (P_T1, P_T2) was noted after they were submerged at 25 °C for two hours. Calculating the maximum load (R_T1, R_T2) required the use of Equations (3) and (4). Equation (5) was used to calculate the ratio of freeze—thaw splitting strength (TSR).

$$R_{T1}=\frac{0.006287P_{T1}}{h_1}$$

(3)

$$R_{T2}=\frac{0.006287P_{T2}}{h_2}$$

(4)

$$TSR=\frac{{\overline{R}}_{T2}}{{\overline{R}}_{T1}}\times 100$$

(5)

2.3.4. Fuzzy Comprehensive Evaluation Analysis Method

(1)

Establish fuzzy evaluation matrix

The scheme set consists of i schemes and j indexes, as shown in Equation (6).

$$R=\left(\begin{array}{ccc}{x}_{11}& \dots & x_{1j}\\ ⋮& \ddots & ⋮\\ x_{i1}& \cdots & x_{ij}\end{array}\right)$$

(6)

In order to eliminate the influence of dimension and numerical range, matrix R is standardized. When x_ij is a positive index, it is calculated according to Equation (7).

$$r_{ij}=\frac{x_{ij}}{x_{i\max }}$$

(7)

When x_ij is a negative index, it is calculated by Equation (8).

$$r_{ij}=1-\frac{x_{ij}}{x_{i\max }}$$

(8)

where x_ij is the element in R matrix. x_imax is the maximum value of the j index. Finally, the membership degree matrix R_p is obtained, as shown in Equation (9).

$$R_p=\left(\begin{array}{ccc}{r}_{11}& \dots & r_{1j}\\ ⋮& \ddots & ⋮\\ r_{i1}& \cdots & r_{ij}\end{array}\right)$$

(9)

(2)

Comprehensive evaluation based on fuzzy mathematics

The weight vector of each index determined by the entropy weight method and the fuzzy evaluation matrix R_p determined by the fuzzy mathematics theory can calculate the relative weights of all alternatives, as shown in Equation (10).

$$L=R_p{W}^{-1}=\left(\begin{array}{ccc}{r}_{11}& \dots & r_{1j}\\ ⋮& \ddots & ⋮\\ r_{i1}& \cdots & r_{ij}\end{array}\right){\left(w_1,w_2,\dots ,w_j\right)}^{-1}=(l_1,l_2,\dots ,l_i)$$

(10)

where w is the weight and l is the comprehensive evaluation index.

(3)

Determining weight value based on entropy weight method

The number of evaluation programs, n, and the evaluation index, m, are used to generate the data matrix, X = (x_ij)_nm, where i = 1, 2,…, n and j = 1, 2,…, m. The initial matrix looks like this:

$$X=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \dots & x_{1m}\\ x_{21}& x_{22}& \dots & x_{2m}\\ \dots & \dots & \dots & \dots \\ x_{n1}& x_{n2}& \dots & x_{nm}\end{array}\right]$$

(11)

Prior to measurement, the data must undergo normalization or dimensionless processing to eliminate discrepancies in the data’s units and other characteristics and to make them comparable, as shown in Equation (12).

$${\mathrm{y}}_{\mathrm{ij}}=\frac{x_{\mathrm{ij}}}{{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{x}_{\mathrm{ij}}}}$$

(12)

Entropy value (H_j) is as shown in Equation (13).

$${\mathrm{H}}_{\mathrm{j}}=-\frac{1}{\ln (\mathrm{m})}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}{y}_{\mathrm{ij}}\ln y_{\mathrm{ij}}}$$

(13)

Coefficient of variation (G_j) is as shown in Equation (14).

$${\mathrm{G}}_{\mathrm{j}}{=1-\mathrm{H}}_{\mathrm{j}}$$

(14)

Weight (ω_j) is as shown in Equation (15).

$${\omega }_{\mathrm{j}}=\frac{{\mathrm{G}}_{\mathrm{j}}}{{\displaystyle \sum _{\mathrm{j}=1}^{\mathrm{m}}{\mathrm{G}}_{\mathrm{j}}}}$$

(15)

2.3.5. Random Forest Analysis Method

The random forest model is an integrated model that generates multiple sets of decision trees jointly for prediction by using a subset of training samples [24]. Each decision tree is generated independently without any pruning, and each node is segmented using a user-defined number of features, chosen randomly. The random forest technique may also provide variable importance assessments to evaluate the intensity of each independent variable’s influence (degree of relevance) on the anticipated dependent variable.

Random forest analysis (RFA) grows and combines plentiful decision trees to construct a powerful “forest”. The most intuitive benefit of RFA is its capability to correct for ordinary regression’s habit of overfitting. When we input a new unlabeled datum, it will be evaluated against all the decision trees created in the set and each tree votes for a class member. The category with the most votes will be the final selected category, and the random forest concept diagram is shown in Figure 2.

3. Results and Discussion

3.1. Aggregate Properties

3.1.1. Morphology Characterization

In Figure 3, SEM images of diabase, granite, basalt, and type A and type B limestone are displayed. Figure 3 shows the morphologies of the five aggregate particles through a 2.36 mm sieve with 500× magnification. This microstructure indicates that the shape of the aggregate profile on the aggregate surface is irregular and not uniformly distributed. This may lead to a different adhesion strength at the asphalt binder—aggregate system interface. At the same time, the asphalt binder film will be most likely to spall from the non-textured parts of the aggregate surface under adverse factors such as loading and flooding.

The SEM image needs to be binarized in order to acquire the aggregate surface’s textural features with accuracy. The grayscale values of the pixel points on the image were set to 0 or 255, thus presenting the whole image with a distinct visual effect of only black and white. To achieve this process, binarization was performed by ImageJ V1.8.0.112 software, and the processed image is shown in Figure 4. After binarization of the SEM images of the five aggregates, it can be found that the variability of micro-texture shapes and other shapes on the surface of the aggregates is large and extremely irregular. In addition, it can be guessed that the morphological characteristics must cause the bonding ability between the asphalt binder and the aggregates, so the microscopic morphological indicators need to be analyzed.

First, the Fractal count function in the ImageJ software was used to calculate the microscopic parameters, and two surface morphological indicators, the roughness and shape factor, were selected. The box dimension method was used to determine the fractal dimension of the binarized picture of the aggregate surface texture. The ImageJ software offers a variety of methods for doing this. It can be seen from Figure 5 that the roughness and shape factors of the five aggregates calculated by the ImageJ software do not differ much. The ranking of the fractal dimension of the surface texture of the five aggregates is: type B limestone, granite, type A limestone, diabase, and basalt. Type B limestone aggregate samples have the most developed surface texture, which is conducive to enhancing the bond between asphalt binder and aggregate when other conditions are the same. In addition, the fractal dimension of granite is larger than that of diabase and basalt and similar to that of limestone. This indicates that the surface roughness of the aggregates in this study is not directly related to their lithology. Secondly, the results of the fractal dimension calculation are related to the fractal dimension calculation method, and different results are obtained by different fractal dimension methods. Therefore, the same selection area and threshold selection method should be employed when using ImageJ software.

3.1.2. Chemical Compositions

Figure 6 displays the chemical compositions of type A limestone, type B limestone, diabase, granite, and basalt as determined by XRF analysis. The acid-base property of aggregate is one of the key factors affecting the adhesion of an asphalt binder—aggregate system. Magmatic rock’s acid-base characteristics are determined by its SiO₂ concentration. A rock is classified as acidic if its silicon dioxide content is greater than 65% and neutral if it is between 52% and 65%. Basic rock and ultrabasic rock have silica contents between 45% and 52%, respectively. Diabase and basalt are hence neutral rocks, whereas granite is an acidic rock. According to the criteria listed above, limestone from sedimentary rock cannot be classified [25]. However, because it contains a lot of CaO, limestone is a basic rock. The chemical compositions and phase compositions of rock aggregates are completely random due to the environmental influence during rock formation. The chemical composition of the same rock varies from one origin to another and from one layer to another. However, the composition of the main minerals of the same rock is generally the same. Therefore, the analytical results of XRD are used in combination with the XRF results.

3.1.3. Phase Compositions

The XRD analysis results of type A limestone, type B limestone, diabase, granite, and basalt are shown in Figure 7. The analysis results show that the type A limestone and type B limestone are mainly composed of two minerals, quartz and dolomite. This indicates that the mineral composition of the two types of limestone is similar, and the mineral element composition is relatively simple. The other three types of rocks are magmatic rocks with more mineral phases, among which basalt contains the most complex mineral phases.

In the system of evaluating the acidity and alkalinity of magmatic rock aggregates, the SiO₂ content is generally considered to be one of the important influencing factors. The main component of mineral phase quartz is SiO₂. The proportion of mineral phase quartz in the five rocks from high to low is granite, type B limestone, type A limestone, diabase, and basalt. The quartz content in diabase, granite, and basalt is 11.51%, 30.1%, and 5.38%, respectively.

By combining XRF and XRD tests on the five types of aggregates, it was found that the SiO₂ content varied. The SiO₂ content of limestone was around 26%, where the element Si was present in the quartz mineral phase and belonged to alkaline dolomitic limestone. The SiO₂ content of diabase and basalt is 55.94% and 54.40%, respectively, where the element Si exists in the ores of silica, feldspar, pyroxene, and hornblende, which are neutral magmatic rocks. The SiO₂ content of granite is 67.23%, and the Si element exists in four minerals, namely chlorite, silica, plagioclase, and black mica, which are acidic deep magmatic rocks.

3.2. Moisture Susceptibility

3.2.1. Modified Boiling Water Test

The asphalt binder is shown peeling off the surfaces of the five aggregates after the boiling water test. By using engineering knowledge and visual inspection, it was discovered that the granite’s peeling was the worst and that the asphalt at the angles of basalt and limestone showed only minor peeling. Diabase and granite’s surface asphalt also showed significant peeling. Granite and asphalt binder clearly adhered the least well, according to the data. To evaluate the outcomes of the boiling water test and verify that the data were visible and quantifiable, the index of peeling rate was used in this test. Figure 8 demonstrates that, with limestone adhesion being superior to granite adhesion, the five aggregates’ peeling rates and the adhesion work of the asphalt binder—aggregate system were roughly on par. Through XRF elemental analysis, this is related to the acidity and alkalinity of the aggregate, where limestone is an alkaline aggregate, about 50% of the composition is CaCO₃, so the surface of the stone is positively charged and reacts chemically with the active material with negative charge in the asphalt binder to form the bond strength, while granite is an acidic aggregate, 60% of the composition is SiO₂, and does not have a strong chemical reaction with the asphalt binder material, so under certain conditions of physical adsorption, the chemical reaction produced by limestone can complement and enhance the bonding properties with asphalt binder.

3.2.2. Immersion Marshall Test

Figure 9 adequately shows which types of aggregates—type A limestone, type B limestone, basalt, diabase, and granite—are used to make the five asphalt mixtures, which are classified according to their water susceptibility. This is consistent with the residual susceptibility of water immersion, MS₀. Figure 10 demonstrates the asphalt mixture formed of granite exhibited excellent susceptibility after 30 min in a water bath at 60 °C. However, the susceptibility of the granite-based asphalt mixture dramatically decreased after 48 h in a water bath at 60 °C. When combined with the MS₀ study in Figure 9, the asphalt mixture formed of granite showed poor water susceptibility.

3.2.3. Freeze—Thaw Splitting Test

As shown in Figure 10, the freeze—thaw splitting strength ratios TSR of asphalt mixtures composed of five aggregates are ranked in order of magnitude as type B limestone, type A limestone, basalt, diabase, and granite. In comparison to basalt and diabase asphalt mixtures, limestone-based asphalt mixtures showed greater MS₀ and TSR values, with no appreciable difference between the two values. Combining Figure 9 and Figure 10, it can be seen that the outcomes, despite the low TSR, were very similar to those of the immersion Marshall test. This result is due to the vacuum saturation process that the freeze—thaw splitting test goes through to enable moisture to enter the spaces in the asphalt mixture. During the freeze—thaw cycle, water expands as it freezes, increasing its volume. The voids in the asphalt mixture may enlarge as a result of this process, making it more vulnerable to harm.

3.3. Statistical Analysis

3.3.1. Fuzzy Comprehensive Evaluation of Water Damage Resistance

Using the fuzzy comprehensive evaluation analysis approach, the results of the research on macroscopic adhesion and the susceptibility of asphalt mixes to moisture were pooled and evaluated. Prior to separating the findings of the positive and negative correlation data, the macro index data were first standardized. The weight coefficients for each index were then determined using the entropy weight method’s calculation procedures (see

Table 3. Normalized results of water damage resistance index data.

Water Damage Resistance Index	Aggregate					ωj
	Type B Limestone	Granite	Basalt	Type A Limestone	Diabase
Wb	0.5101	0.3348	0.9638	1.0000	1.0000	0.1809
MS	1.0000	1.0000	1.0000	0.9869	0.9232	0.1962
MS0	0.0000	0.3240	0.6503	0.7240	0.8598	0.1756
RT1	0.9008	0.0000	0.0000	0.0000	0.0000	0.2623
TSR	0.8320	0.7056	0.8150	0.9753	0.9185	0.185

). The fuzzy evaluation matrix R was established based on the results of macroscopic adhesion tests and moisture susceptibility tests of the five aggregates, and the matrix R was standardized to eliminate the effects of magnitude and quantitative extremes to finally obtain the affiliation matrix R_p. Finally, the fuzzy mathematical model was used for comprehensive evaluation, and the evaluation results are shown in

Table 4. Results of fuzzy comprehensive evaluation of water damage resistance.

Comprehensive Evaluation Index	Aggregate
	Type B Limestone	Granite	Basalt	Type A Limestone	Diabase
li	0.940377	0.568443	0.886701	0.856893	0.748623

.

The results of the fuzzy comprehensive evaluation analysis of the macroscopic adhesion index and the moisture susceptibility index show that the type B limestone, basalt, type A limestone, diabase, and granite had the highest scores among the five aggregates. Notably, granite received the lowest rating, while basalt and limestone received the highest ratings based on the results of each sub-index. This general consistency between the rankings based on li and the results of each sub-index is remarkable.

3.3.2. Correlation Analysis of Aggregate Characteristics and Water Damage Resistance

To conduct the correlation study of aggregate characteristics and water damage resistance, the output layer of the random forest analysis model was the complete evaluation index l_i of the asphalt mixture against water damage. The random forest analysis model was first created using the RStudio 4.2.1 program. The output layer’s resistance to water damage (l_i) of the asphalt mixture was then made the dependent variable, while other aggregate characteristic parameters were employed as covariates. Figure 11 displays the independent variable importance analysis. Among the indicators of aggregate characteristics, the degree of correlation with water damage resistance is in the order of SiO₂, roughness, CaO, shape factor, fractal dimension, and quartz. The CaO content has an effect on the ability to resist water damage, but is not the dominant effect among chemical elements, a conclusion shared in previous scholarly studies. Roughness, shape factor, and fractal dimension are relatively well correlated with water damage resistance, with roughness taking the lead as a reference.

4. Conclusions

In this research, the morphology characterization, chemical composition, and phase composition of the five aggregates were first studied. The five asphalt binder—aggregate systems’ water susceptibility was assessed using the engineering approach for assessing the resistance of asphalt mixtures to water damage. Their pertinent characteristic parameters were determined. Then the water susceptibility of the five asphalt binder—aggregate systems was evaluated comprehensively by using the fuzzy comprehensive evaluation method. Finally, sensitivity analysis of the aggregate characteristic parameters was carried out by a random forest analysis model, so as to achieve the optimal screening of the aggregate characteristic parameters.

(1)

During sensitivity analysis of each parameter of the aggregate properties using the random forest analysis, the SiO₂ content of the aggregate had the highest importance, and the roughness had the highest importance for the morphology characterization. The water susceptibility of the asphalt binder—aggregate system could be expressed by the SiO₂ content and roughness of the aggregate characteristic parameters.

(2)

The morphology characterization of the aggregates was analyzed by binarizing the SEM photographs of different aggregate materials, and it was found that the surface texture of the limestone samples was the most developed. Among them, the fractal dimension of granite was similar to that of limestone, which indicated that the surface roughness of the aggregates in this study was not directly related to their lithology.

(3)

The chemical element and compound composition combination of the aggregates was completely random due to the environmental influence during rock formation, and the chemical composition of the same rock varied for different places of origin and layers. However, the composition of the main minerals of the same rock was largely the same.

(4)

Using different aggregate compositions and the fuzzy comprehensive evaluation analysis approach, the water susceptibility index of the asphalt binder—aggregate system was investigated. The results showed that, whereas limestone had the best resilience to water damage, granite had the weakest.