Machine Learning-Enhanced Nanoindentation for Characterizing Micromechanical Properties and Mineral Control Mechanisms of Conglomerate

Abstract

Conglomerate reservoirs present significant technical challenges during drilling operations due to their complex mineral composition and heterogeneous characteristics, yet the quantitative relationships between mineral composition and microscopic mechanical behavior remain poorly understood. To elucidate the variation patterns of conglomerate micromechanical properties and their mineralogical control mechanisms, this study develops a novel multi-scale characterization methodology. This approach uniquely couples nanoindentation technology, micro-zone X-ray diffraction analysis, and machine learning algorithms to systematically investigate micromechanical properties of conglomerate samples from different regions. Hierarchical clustering algorithms successfully classified conglomerate micro-regions into three lithofacies categories with distinct mechanical differences: hard (elastic modulus: 81.90 GPa, hardness: 7.83 GPa), medium-hard (elastic modulus: 54.97 GPa, hardness: 3.87 GPa), and soft lithofacies (elastic modulus: 25.21 GPa, hardness: 1.15 GPa). Correlation analysis reveals that quartz (SiO₂) content shows significant positive correlation with elastic modulus (r = 0.52) and hardness (r = 0.51), while clay minerals (r = −0.37) and plagioclase content (r = −0.48) exhibit negative correlations with elastic modulus. Mineral phase spatial distribution patterns control the heterogeneous characteristics of conglomerate micromechanical properties. Additionally, a random forest regression model successfully predicts mineral content based on hardness and elastic modulus measurements with high accuracy. These findings bridge the gap between microscopic mineral properties and macroscopic drilling performance, enabling real-time formation strength assessment and providing scientific foundation for optimizing drilling strategies in heterogeneous conglomerate formations.

1. Introduction

Conglomerate, as a typical terrigenous clastic rock, exhibits complex composition and structural characteristics that result in variable properties, high wear resistance, and poor drillability [1,2,3]. These characteristics seriously affect the safety and efficiency of drilling operations, presenting significant challenges to drilling efficiency [2,4,5]. Tight conglomerate reservoirs comprise gravels and matrix, making rock mechanical property characterization difficult and preventing effective guidance for drilling and hydraulic fracturing operations in tight conglomerate oil and gas reservoirs [6]. For example, in Norway’s North Sea Luno oil field, drilling rates in conglomerate reservoirs are only 0.5–3.5 m/h, with rates typically maintained between 1–2 m/h, severely constraining economic development benefits [7], while China’s Mahu oil field faces serious hydraulic fracturing challenges including fracture initiation and proppant transport due to its complex structural variations [8,9]. Therefore, developing an in-depth understanding of conglomerate mechanical properties and fracturing mechanisms has important theoretical significance and practical value for improving drilling efficiency in conglomerate formations [10,11].

Existing research demonstrates that conglomerate mechanical performance is determined by microscopic mineral composition and structure, with gravel content being the key factor controlling conglomerate mechanical performance [12,13,14]. When gravel content falls below a critical threshold, the rock undergoes brittle–plastic transition, and macroscopic mechanical performance becomes primarily controlled by cementing materials. Therefore, clarifying the microscopic strength characteristics of conglomerate cement and gravels is crucial for guiding drilling operations in conglomerate formations [12,13]. However, traditional macroscopic mechanical testing methods cannot accurately capture microscopic mechanical differences among internal components in conglomerates [15]. Nanoindentation technology, as a method for determining mechanical properties of both homogeneous and heterogeneous materials, can precisely measure mechanical parameters such as elastic modulus and hardness at the micro-nanometer scale. This provides a new approach for studying conglomerate microscopic mechanical characteristics and is essential for understanding conglomerate microscopic strength features [16,17,18]. Discrete element numerical simulation results indicate that cement strength and gravel characteristics significantly influence conglomerate heterogeneity and macroscopic mechanical performance [15]. X-ray diffraction technology, as an important tool for mineral characterization, has been widely applied in mineral identification and quantitative analysis in earth sciences. This technology determines mineral phase composition in samples by analyzing characteristic diffraction patterns produced by X-ray interactions with crystal structures [19]. Therefore, this paper characterizes and predicts microscopic strength characteristics of conglomerates using nanoindentation methodology. The Oliver–Pharr analytical method is an indentation experimental analysis approach based on elastic contact theory [20]. It effectively characterizes material mechanical properties such as elastic modulus and hardness [21], establishing a theoretical foundation for quantitative evaluation of mechanical properties across different mineral phases. In recent years, scholars have successfully characterized mechanical property differences among major rock-forming minerals in rocks such as shale and granite using nanoindentation technology [22,23]. Shale microscopic mechanical property research demonstrates that combining nanoindentation with numerical simulation methods can effectively evaluate shale mechanical properties under varying felsic and clay mineral content [24,25,26]. Additionally, research on three major granite minerals—quartz (SiO₂), feldspar, and mica—at different indentation depths revealed that quartz and feldspar exhibit increasing crack length with increasing indentation depth, while mica shows minimal radial cracking [27], providing important reference for multi-scale rock mechanics research. Furthermore, random forest algorithms, as ensemble learning methods, improve model accuracy and stability by constructing multiple decision trees and integrating their prediction results. These algorithms have achieved excellent application results in mineral exploration prediction, rock burst prediction, and rock strength prediction [28,29]. Deep learning and machine learning technologies also demonstrate tremendous potential in mineral microscopic image recognition, with integrated multiple machine learning algorithms significantly improving rock mineral identification accuracy [30]. Therefore, developing multi-scale characterization methods that integrate nanoindentation, micro-area X-ray diffraction, and machine learning algorithms is highly significant. These methods can systematically reveal mechanical behavior patterns of conglomerate microscopic components and guide efficient rock breaking technology development in conglomerate formations.

In summary, existing research primarily focuses on macroscopic-scale conglomerate mechanical property testing. Research on microscopic mechanical characterization of conglomerates as complex multi-phase systems remains limited and lacks systematic comparative analysis of gravel and matrix mechanical properties, as well as quantitative relationship models between mineral composition and mechanical properties. Understanding of how mineral components influence local mechanical properties at the microscopic scale is insufficient. Additionally, while traditional X-ray diffraction technology can identify mineral composition, establishing direct correspondence with nanoindentation mechanical parameters remains challenging. To construct quantitative relationships between mineral composition and mechanical properties, this paper utilizes real cores from different stratigraphic levels, combining nanoindentation experiments, micro-area X-ray diffraction analysis, unsupervised clustering algorithms, and random forest prediction models. This approach systematically reveals heterogeneous distribution patterns of conglomerate microscopic mechanical characteristics and their mineralogical control mechanisms, establishes intelligent prediction models from mechanical parameters to mineral composition, and identifies microscopic strength characteristics of conglomerates across different blocks.

2. Materials and Methods

2.1. Experimental Equipment

Nanoindentation experiments were performed using a KLA G200 nanoindenter for hardness and elastic modulus testing, as shown in Figure 1a. The instrument features excellent technical specifications: displacement resolution < 0.01 nm, maximum indentation depth > 500 μm, test load range of 10 μN to 500 mN, load resolution ≤ 1 nN, displacement noise floor < 0.05 nm, and instrument stiffness > 5 × 10⁶ N/m. The testing area covers 100 mm × 100 mm with XY positioning accuracy of 1 μm. The equipment includes 10× and 40× objective optical microscope systems and an optional atomic force microscope module. It supports multiple testing modes including hardness and modulus testing, continuous stiffness measurement (CSM), dynamic mechanical analysis (DMA), and nano-scratching [19].

Considering the strong heterogeneous characteristics of conglomerate, this study employed single stiffness measurement (SSM) instead of continuous stiffness measurement (CSM). Beyond noise reduction, SSM was selected due to its superior performance in heterogeneous materials, providing precise depth control to avoid crossing mineral boundaries and eliminating oscillation-induced artifacts at phase interfaces, which is essential for accurate mechanical property determination in multi-phase conglomerate systems. The experiment utilized a Berkovich diamond indenter, as shown in Figure 1b, with tip roundness < 20 nm, making it suitable for detecting mechanical properties of various minerals in complex materials.

Micro-area X-ray diffraction analysis employed a D8 Advance X-ray diffractometer (Bruker, Karlsruhe, Germany), as shown in Figure 1c, for precise identification of mineral phase composition in conglomerate nanoindentation regions. The equipment features a ceramic copper target X-ray tube (focal spot 0.4 × 12 mm), maximum output power of 3 kW, operating voltage of 40 kV, and current of 40 mA, utilizing Cu Kα₁ radiation (λ = 0.15406 nm) as the incident radiation source. The diffractometer employs a θ/θ vertical goniometer configuration with a 2θ scanning range of −110° to 168°, angular resolution of 0.0001°, and scanning step size of 0.02°, optimizing both data quality and acquisition efficiency. The system incorporates a high-sensitivity two-dimensional area detector (maximum count rate of 1 × 10⁹ cps) that effectively captures weak diffraction signals. Current and voltage stability exceeds ±0.005% (under 10% external voltage fluctuation), ensuring stable long-term data acquisition. Standard samples were used for calibration prior to experiments, guaranteeing diffraction peak angle deviation within ±0.01°.

2.2. Experimental Material

Samples were extracted from conglomerate cores from five wells and embedded using epoxy resin (EpoFix, Struers, Copenhagen, Denmark) under vacuum conditions (−0.8 bar for 15 min) to eliminate air bubbles. The curing process involved room-temperature curing (23 ± 2 °C) for 8 h followed by post-curing at 40 °C for 2 h. Specimens were prepared as standard cylinders with 30 mm diameter and 10 mm height, then subjected to progressive polishing using SiC papers (240, 400, 800, 1200, 2400 grit) followed by diamond paste polishing (6 μm, 3 μm, 1 μm) to achieve surface roughness < 0.1 μm. Prior to testing, reference points were marked on sample surfaces, and indentation positions were determined to provide precise positioning references for subsequent micro-area X-ray diffraction experiments. This study employed the single stiffness method based on Oliver–Pharr theory, which calculates contact stiffness by analyzing the initial slope of unloading curves to obtain material hardness and elastic modulus. This method is well-suited for heterogeneous materials such as conglomerate.

Considering conglomerate heterogeneity, different color markings were used to identify test regions: red indentation points marked gravel areas, yellow indentation points marked cement areas, and blue indentation points marked matrix areas, as shown in Figure 2. A total of 210 test points were established. Subsequently, micro-area X-ray diffraction testing was conducted to build a mineral composition database for indentation areas, enabling correlation analysis between mechanical properties and mineral composition. To ensure adequate representation of conglomerate heterogeneity, the distribution of 210 test points was determined based on quantitative analysis of component proportions in each sample. Test points were systematically allocated using a color-coded classification system (red for gravels, blue for matrix, yellow for cement) with point density proportional to the measured areal percentages of each component observed under optical microscopy.

2.3. Clustering Analysis Methods

Clustering analysis, as an unsupervised learning method applied to rock mechanical property analysis, can automatically identify and group samples with similar characteristics. Unlike traditional rock classification methods, clustering analysis requires no predefined category labels but discovers potential grouping patterns through the intrinsic structure and similarity patterns within the data itself. This approach is significant for revealing complex microscopic mechanical property distribution patterns in conglomerate. In highly heterogeneous material systems such as conglomerate, mechanical properties may vary significantly across different regions, and traditional empirical classification methods often cannot accurately capture such complex variational characteristics. Through objective mathematical algorithms, clustering analysis can automatically identify regions with similar mechanical properties based on elastic modulus and hardness data obtained from nanoindentation tests, thereby providing a scientific basis for lithofacies classification of conglomerate. This study employed two main clustering techniques to analyze rock mechanical properties: K-means clustering and hierarchical clustering. The most suitable clustering strategy for conglomerate mechanical property analysis was determined through comparative analysis of both methods’ classification performance.

The K-means clustering algorithm uses distance metrics to assign data points to K predefined clusters through an iterative optimization process [31,32]. The core objective is to minimize the sum of squared distances from each data point to its respective cluster center. The objective function can be expressed as:

$$J={\displaystyle \sum _{j=1}^k{\displaystyle \sum _{i=1}^n\Vert x_i^{(j)}-c_j{\Vert }^2}}$$

(1)

where $x_i^{\left(j\right)}$ represents the i-th data point belonging to the j-th cluster, $c_j$ is the center of the j-th cluster, and ${\left|x_i^{\left(j\right)}-c_j\right|}^2$ represents the squared Euclidean distance. In the K-means algorithm’s iterative process, K data points are first randomly selected as initial cluster centers. Then, the distance from each data point to each center is calculated, and each point is assigned to the cluster with the nearest center. After assignment, cluster centers are updated by calculating the mean of all points within each cluster. These two steps—“assign data points” and “update cluster centers”—are repeatedly executed until cluster division stabilizes or the preset maximum iteration number is reached.

Hierarchical clustering constructs data hierarchical structures by evaluating inter-sample similarity using bottom-up (agglomerative) or top-down (divisive) approaches [33]. This study selected agglomerative hierarchical clustering based on the Ward linkage method, which aims to minimize intra-cluster variance increment at each merging step. The Ward method specifically defines a distance metric criterion: the error sum of squares (ESS) increment before and after merging two clusters:

$$d_{A,B}=\Delta ESS=ES{S}_{A\cup B}-ES{S}_A+ES{S}_B$$

(2)

$$ES{S}_A={\displaystyle \sum _{i\in A}\Vert x_i-\overline{x_A}{\Vert }^2}$$

(3)

where $ES{S}_A$ represents the error sum of squares for cluster A, and $\overline{x_A}$ is the centroid of cluster A. The Ward method favors merging clusters that produce the minimum $ESS$ increment after merging, thus generating compact, spherical clusters.

This study first performed standardization preprocessing on collected nanoindentation test data points to eliminate comparison bias between elastic modulus and hardness due to dimensional differences:

$$z_i={\displaystyle \frac{x_i-\mu }{\sigma }}$$

(4)

where $z_i$ is the standardized value, $x_i$ is the original value, $\mu$ is the mean, and $\sigma$ is the standard deviation. Both K-means and hierarchical clustering methods were applied for classification, combined with techniques such as the elbow method and dendrograms to determine optimal cluster numbers and evaluate the classification performance of both methods.

2.4. Random Forest Prediction Method

Random forest regression, as an ensemble learning method, enhances model prediction accuracy and stability by constructing multiple independent decision trees and integrating their prediction results. Its theoretical foundation is based on Bootstrap resampling, which constructs diverse decision trees through random sampling with replacement from the original dataset while introducing random feature subset selection during node splitting. This ultimately forms robust prediction models through mean aggregation. This design, based on diversity and ensemble principles, reduces prediction variance while maintaining moderate model bias, making it well-suited for prediction tasks in rock mechanics.

Model construction was implemented using Python (version 3.13.5) with the Scikit-learn library (version 1.6.1), building 100 decision trees with optimized feature sampling strategies. Elastic modulus and hardness measured by nanoindentation served as input features to predict quartz and clay mineral content in micro-areas. The algorithm’s core workflow comprises four steps: First, Bootstrap sampling randomly selects N samples with replacement from the original training dataset to form new training sets. Second, during each node split, m feature subsets are randomly selected from all available features. Third, decision trees are constructed based on the selected feature subsets, using Gini impurity or mean squared error as the splitting criterion. Finally, for regression problems, the final prediction result represents the average of all decision tree predictions.

Model performance was comprehensively evaluated using multiple metrics. The coefficient of determination (R²) measures the model’s ability to explain data variation, root mean square error (RMSE) evaluates deviation between predicted and actual values, and normalized root mean square error (NRMSE) eliminates dimensional effects.

3. Results

3.1. Nanoindentation

The nanoindentation experiments comprised 210 measurement points, yielding 208 valid data points. To characterize the heterogeneity of conglomerate microscopic mechanical properties, results were categorized into three groups based on mechanical response parameters, with their corresponding load–displacement curves plotted in Figure 3. The first category exhibits elastic modulus exceeding 75 GPa and hardness values greater than 8 GPa, with load–displacement curves shown in Figure 3a,b. The second category shows elastic modulus between 35–70 GPa and hardness values ranging from 4–8 GPa, as illustrated in Figure 3c,d. The third category displays elastic modulus between 0–35 GPa and hardness below 4 GPa, as shown in Figure 3e,f.

Figure 4 shows the frequency distribution of elastic modulus and hardness categories. Analysis indicates that when classified by elastic modulus, Class 2 (medium elastic modulus) indentation points predominate (110 points), followed by Class 1 (high elastic modulus) indentation points (53 points), while Class 3 (low elastic modulus) indentation points are fewest (46 points). However, when hardness serves as the classification criterion, a different distribution trend emerges: Class 3 (low hardness) exhibits the highest frequency, Class 2 (medium hardness) ranks second, and Class 1 (high hardness) shows the lowest frequency. This distribution difference reveals the heterogeneous characteristics of conglomerate components across different mechanical parameters, providing a foundation for subsequently developing conglomerate strength classification models based on mineral composition.

3.2. Micro-Area XRD

To investigate the relationships between elastic modulus, hardness, and mineral components, micro-area X-ray diffraction experiments were conducted by analyzing mineral distribution data at corresponding indentation positions. Indentation points with different elastic moduli were classified according to major mineral components (clay minerals, feldspar, and quartz).

Figure 5 shows the micro-area XRD patterns of Class 3 indentation points, while Appendix A presents the mineral composition data of Class 3 low elastic modulus indentation points. Class 3 indentation point samples exhibit significantly lower mechanical properties, with average elastic modulus of 22.71 GPa and hardness of only 0.75 GPa, representing merely 25% and 9% of the corresponding Class 1 indentation point values, respectively. The mineralogical characteristics of Class 3 indentation points show: lowest quartz content (36.60%), highest plagioclase ((Na,Ca)(Al,Si)₄O₈) content (32.33%), and highest clay mineral content (19.60%).

The fundamental causes for reduced mechanical properties in Class 3 indentation points can be attributed to three factors: high clay mineral content creates numerous planes of structural weakness and potential slip surfaces; elevated plagioclase proportions compromise the rock’s microscopic structural integrity; and insufficient quartz content weakens the structural support framework, resulting in the low-strength characteristics observed in Class 3 indentation points.

As shown in Figure 6, the micro-area XRD patterns of Class 2 indentation points reflect their mineral composition characteristics. The Class 2 indentation point mineral composition data in Appendix A indicate that this lithofacies exhibits mechanical properties intermediate between Class 3 (soft) and Class 1 indentation points, with average elastic modulus of 52.78 GPa and hardness of 3.58 GPa, representing approximately 60% and 42% of Class 1 indentation point values, respectively.

In terms of mineral composition, Class 2 indentation points exhibit the following characteristics: moderate quartz content (42.84%), relatively high plagioclase content (22.13%), highest calcite (CaCO₃) content (8.09%), and clay mineral content (18.21%) significantly higher than Class 1 indentation points. This phenomenon indicates that when quartz content is insufficient to provide adequate structural support, the presence of plagioclase and clay minerals significantly reduces conglomerate mechanical performance. Additionally, Class 2 indentation points contain the highest calcite content. Considering calcite’s low hardness (Mohs hardness 3) and tendency to fracture along cleavage planes, this component increases rock brittleness and consequently reduces overall mechanical strength.

As shown in Figure 7, the micro-area XRD patterns of Class 1 indentation points reveal their mineral composition characteristics. The Class 1 indentation point mineral composition data in Appendix A indicate that Class 1 indentation point samples exhibit high-strength mechanical characteristics, with an average elastic modulus and hardness of 90.85 GPa and 8.42 GPa, respectively. Mineral composition analysis shows that Class 1 indentation points display significant quartz enrichment (average content 59.43%), moderate plagioclase content (17.53%), and the lowest clay mineral content (10.63%). As a high-hardness mineral with Mohs hardness of 7, the high quartz content directly enhances the overall strength performance of Class 1 indentation points. Class 1 indentation point samples exhibit more balanced mineral composition proportions and tighter inter-particle contacts, effectively reducing microcrack occurrence probability while improving the rock’s resistance to deformation and overall strength performance.

Through comparative analysis of the three lithofacies, correlation patterns between gravel and matrix mechanical properties and their mineral composition were revealed. Research demonstrates that quartz content serves as the key indicator for distinguishing lithofacies categories, exhibiting a clear gradient distribution: Class 1 indentation points, 59.43%; Class 2 indentation points, 42.84%; and Class 3 indentation points, 36.60%. This distribution is highly consistent with the decreasing trend in mechanical properties. In contrast to quartz content changes, plagioclase and clay mineral contents increase from Class 1 to Class 3, rising from 17.53% to 32.33% and from 10.63% to 19.60%, respectively. The content variations of these three major minerals constitute the fundamental mechanism affecting conglomerate microscopic mechanical properties.

The influence mechanisms of mineral composition exhibit significant differences among different indentation point classes: high quartz content in Class 1 indentation points enhances overall rock strength; in Class 2 indentation points, increased plagioclase and clay content reduces rock strength; in Class 3 indentation points, the combined weakening effects of clay minerals and plagioclase exceed the strengthening effect of quartz. This combined weakening effect dominates the mechanical performance, resulting in significantly reduced overall rock mechanical strength.

3.3. Clustering Analysis

Clustering analysis, as an unsupervised machine learning method, enables objective grouping based on intrinsic data structural characteristics without prior classification standards, providing a reliable scientific approach for quantitative classification of conglomerate microscopic mechanical properties. This study employed two complementary algorithms—K-means clustering and hierarchical clustering—for comparative verification, successfully dividing 208 nanoindentation test data points into three lithofacies categories with significant differences based on their mechanical response characteristics.

As shown in Figure 8, elastic modulus and hardness parameters obtained from nanoindentation experiments show distinct clustering patterns in two-dimensional feature space. This non-random spatial distribution clearly indicates the presence of multiple microscopic regions with significantly different mechanical properties within conglomerate samples. This pattern demonstrates the control exerted by different mineral particles and their spatial arrangements on local mechanical properties in conglomerate as a composite material system.

Both the Elbow method and Dendrogram analysis verified the statistical validity of the three-cluster division scheme. As shown in Figure 9a, the within-cluster sum of squares (WCSS) curve for K-means clustering exhibits a significant inflection point at k = 3, demonstrating the “elbow” phenomenon. This indicates that the three-cluster configuration achieves optimal balance between minimizing intra-cluster variance and avoiding over-segmentation. The hierarchical clustering dendrogram in Figure 9b reveals three distinct hierarchical branch structures, with fusion nodes between branches located at larger Euclidean distance thresholds, confirming the fundamental differences among lithofacies categories.

Clustering analysis of nanoindentation data indicates that K-means and hierarchical clustering methods yield silhouette coefficients of 0.5039 and 0.4916, respectively, showing no significant difference. Compared to the spherical cluster assumption of K-means clustering, hierarchical clustering algorithms can more accurately identify natural partitioning patterns and irregular boundary features of conglomerate mechanical properties. As shown in Figure 10, hierarchical clustering analysis based on the Ward linkage criterion successfully divided 208 test data points into three lithofacies categories with significant mechanical response differences, exhibiting distinct property gradient distribution characteristics. Hard lithofacies (cluster 1) account for 27.9% of total samples, demonstrating high mechanical strength with mean hardness of 7.83 GPa and mean elastic modulus of 81.90 GPa. Medium-hard lithofacies (cluster 2) comprise the majority of samples at 44.7%, showing moderate mechanical strength with mean hardness of 3.87 GPa and mean elastic modulus of 54.97 GPa. Soft lithofacies (cluster 3) represent 27.4% of total samples, displaying relatively low mechanical strength with mean hardness of only 1.15 GPa and mean elastic modulus of 25.21 GPa.

From Figure 10b, the soft lithofacies region, primarily composed of clay minerals, exhibits an essentially linear relationship between elastic modulus and hardness. Medium-hard lithofacies, due to variable clay mineral and feldspar mixing ratios, deviate from traditional elastic modulus-hardness relationships. In contrast, the hard lithofacies region, enriched in quartz, shows elastic modulus increasing significantly with hardness at a more pronounced rate.

Based on clustering analysis results, the spatial distribution pattern of indentation points is shown in Figure 11. The vast majority of red triangular test points marked as gravel areas correspond to hard lithofacies (red numbers) or medium-hard lithofacies (yellow numbers). However, some wells exhibit relatively low gravel strength, resulting in classification as medium-hard lithofacies. Parts of gravel areas in rock samples 3 and 8 show soft lithofacies characteristics. Combined with drilling rate curve data from related wells, this indicates that drilling rate significantly increases when both gravels and matrix simultaneously display low strength characteristics.

This quantitatively verifies the geological understanding that gravel components generally possess higher mechanical strength. In contrast, blue triangular test points marked as matrix areas mainly exhibit medium-hard to soft lithofacies characteristics. The significant variability in their mechanical properties can be attributed to three factors: 1. boundary effects of test point spatial positioning, where some test points are actually located in gravel-matrix transition zones; 2. limitations of visual determination under optical microscopy, leading to misidentification of micro-areas rich in high-strength minerals as matrix; 3. the inherent strong compositional heterogeneity of the matrix itself.

Comprehensive analysis indicates that individual gravels are relatively homogeneous internally and mostly exhibit high elastic modulus, with only muddy gravels showing moderate mechanical properties. In contrast, matrix areas have complex and variable compositions, containing different proportions of clay minerals, feldspar, and quartz particles. Their weak microscopic bonding forces make them prone to deformation and failure, resulting in relatively low elastic modulus. The diversity of mineral components in conglomerate constitutes the fundamental cause of spatial differences in elastic modulus, while the complex relationships between mineral composition and mechanical properties in the matrix reflect the diversity and complexity of overall conglomerate mechanical behavior.

Based on quantitative analysis data from 17 indentation points with known mineral composition, this study constructed a Pearson correlation coefficient matrix between mineral content and mechanical parameters, as shown in Figure 12, systematically revealing mineralogical control mechanisms of conglomerate micro-area mechanical behavior.

Quartz content shows significant positive correlation with mechanical properties (elastic modulus r = 0.52, hardness r = 0.51, p < 0.01), clearly demonstrating the significant strengthening effect of quartz as a rigid framework mineral on micro-area mechanical strength. This quantitative relationship is highly consistent with quartz’s intrinsic crystal structure characteristics: quartz possesses a highly ordered three-dimensional silica tetrahedra network structure, providing excellent mechanical strength (Mohs hardness 7) and outstanding chemical stability. In conglomerate microstructure, quartz particles form a solid load-bearing framework through dense packing, effectively dispersing stress concentrations and providing strong deformation resistance.

Conversely, clay mineral content exhibits negative correlation with mechanical properties (elastic modulus r = −0.37, hardness r = −0.30), indicating that their unique layered silicate structure and relatively weak interlayer bonding forces are the primary factors causing mechanical property deterioration. In clay mineral layered structures, adjacent structural layers are maintained primarily by weaker van der Waals forces and limited hydrogen bonds. Interlayer bonding energy is significantly lower than intralayer covalent bond strength, making them highly susceptible to interlayer shear slippage under external stress, thereby creating stress concentration zones and preferred failure pathways.

Plagioclase content demonstrates strong negative correlation with mechanical parameters (elastic modulus r = −0.48, hardness r = −0.43), primarily attributed to plagioclase having moderate cleavage characteristics and relatively low hardness values (Mohs hardness 6–6.5). This makes it prone to fracturing along specific crystal faces under stress, forming mechanical weak points.

3.4. Random Forest Prediction Model

Based on mechanical parameter and mineral content paired data from 17 calibrated samples, the well-trained random forest regression model successfully achieved high-precision prediction of mineral composition for the remaining 191 indentation points. Prediction results are shown in Figure 13a,b. Analysis indicates that the three lithofacies categories exhibit clear boundary characteristics in the predicted mineralogical feature space: hard lithofacies (dark blue region) show high quartz content (>50%) and low clay mineral content (<15%); medium-hard lithofacies (light blue region) display transitional characteristics with moderate quartz content (30–50%) and moderate clay mineral content (15–25%); soft lithofacies (brown region) exhibit low quartz content (<30%) and high clay mineral content (>25%).

Figure 13c,d present comparison results between random forest model predictions and measured values. The scatter plot distribution patterns show that the vast majority of predicted data points cluster tightly around the ideal fitting line y = x, clearly demonstrating high consistency between model predictions and actual measured values. Notably, quartz content prediction achieves the highest accuracy in the moderate concentration range (30–60%), while clay mineral content prediction maintains consistently high accuracy across the entire concentration range.

Quantitative performance evaluation results indicate that both mineral content prediction models achieved excellent fitting performance. The coefficient of determination R² for quartz content prediction is 0.8106, while the clay mineral content prediction model reaches R² = 0.8529, indicating that the models effectively explain over 80% of mineral content variation. Pearson correlation coefficients between predicted and measured values reach 0.9436 and 0.9648, respectively, demonstrating strong linear relationships between predictions and actual mineral contents. Additionally, normalized root mean square errors (NRMSE) for both quartz and clay mineral content prediction models are approximately 0.10, meaning prediction errors represent only 10% of the true value variation range. This indicates excellent prediction accuracy and good generalization ability.

4. Discussion

Based on our hierarchical clustering analysis results, we can provide data-driven thresholds for drilling tool selection: (1) Hard lithofacies (mean elastic modulus 81.90 GPa, mean hardness 7.83 GPa): These formations, representing 27.9% of samples, require enhanced cutting technologies and reduced penetration rates; (2) Medium-hard lithofacies (mean elastic modulus 54.97 GPa, mean hardness 3.87 GPa): Comprising 44.7% of samples, these formations allow standard drilling parameters with moderate optimization; (3) Soft lithofacies (mean elastic modulus 25.21 GPa, mean hardness 1.15 GPa): Representing 27.4% of samples, these formations permit aggressive drilling parameters with higher penetration rates. These clustering-derived thresholds provide an objective framework for drilling optimization, though field validation is needed to establish comprehensive operational guidelines.

This study successfully revealed the heterogeneous distribution patterns of micromechanical properties in conglomerate using nanoindentation technology, demonstrating significant advantages over traditional macroscopic testing methods. Traditional uniaxial and triaxial compression tests can only obtain overall mechanical parameters of conglomerate and cannot distinguish microscopic strength differences among gravels, cement, and matrix. In contrast, nanoindentation technology can precisely identify the mechanical response characteristics of different mineral phases at the micro-nanometer scale. However, this technology has certain limitations: (1) Representativeness issues—the testing range of a single indentation point is only hundreds of nanometers, making it difficult to fully represent the overall mechanical behavior of complex conglomerate systems; (2) Sample preparation effects—polishing treatment may damage the microstructure of soft minerals (such as clay minerals), thereby affecting testing accuracy; (3) Testing condition sensitivity—the selection of indentation depth and direction significantly affects the measurement accuracy of mechanical parameters for different minerals, particularly anisotropic minerals such as mica.

The random forest regression model demonstrated excellent performance in mineral content prediction by ensembling multiple decision trees to effectively reduce overfitting risks, achieving coefficients of determination of 0.8106 and 0.8529 for quartz and clay mineral content prediction, respectively. This provides an effective approach for establishing quantitative relationships between mechanical parameters and mineral composition. However, current models still face challenges: the limited training dataset size (only 17 samples) affects generalization ability, and the prediction scope is limited to quartz and clay minerals, with prediction accuracy for other rock-forming minerals such as feldspar and carbonates requiring improvement. Future research should expand the dataset size, introduce deep learning algorithms combined with mineral microscopic image features, construct more comprehensive multi-mineral component prediction models, and achieve dynamic lithology identification and parameter optimization during conglomerate drilling processes.

The true engineering value of the random forest model developed in this study lies in its potential for real-time application during drilling operations. While the current model relies on nanoindentation data obtained in a laboratory setting, a clear pathway exists for its deployment in the field. The key is to establish a bridge between the model’s input parameters (micro-hardness and elastic modulus) and real-time data acquired from downhole sensors, such as those in Measurement-While-Drilling (MWD) or Logging-While-Drilling (LWD) toolsets.

5. Conclusions

Based on nanoindentation technology, micro-area X-ray diffraction analysis, and machine learning algorithms, this study systematically revealed the heterogeneous distribution patterns of conglomerate microscopic mechanical properties and their mineralogical control mechanisms, establishing intelligent prediction models from mechanical parameters to mineral composition. The research results have important theoretical significance and practical value for deepening understanding of rock breaking mechanisms in conglomerate formation drilling and guiding efficient drilling technology development. The main conclusions are as follows:

(1)

A quantitative classification system for conglomerate microscopic mechanical properties was established using clustering analysis. Through 208 effective nanoindentation test data points, conglomerate microscopic components were successfully divided into three lithofacies categories with significant mechanical differences: hard lithofacies, medium-hard lithofacies, and soft lithofacies, with elastic moduli of 81.90 GPa, 54.97 GPa, and 25.21 GPa, respectively, and hardness values of 7.83 GPa, 3.87 GPa, and 1.15 GPa, respectively. This provides a scientific basis for quantitative characterization of conglomerate heterogeneity.

(2)

The control mechanisms of mineral composition on conglomerate microscopic mechanical properties were elucidated. Quartz content exhibits significant positive correlation with both elastic modulus and hardness, with correlation coefficients of 0.52 and 0.51, respectively. Clay mineral and plagioclase contents show negative correlations with mechanical properties. When indentation areas simultaneously cover multiple mineral components, the weakening effects of clay and plagioclase on microscopic strength exceed the strengthening effect of quartz, resulting in micro-areas exhibiting soft rock characteristics. Spatial distribution differences in mineral composition constitute the fundamental cause of strong heterogeneity in conglomerate.

(3)

An intelligent prediction model for mineral content based on mechanical parameters was constructed. Using random forest algorithms, high-precision inversion from nanoindentation mechanical parameters to mineral composition was achieved. Determination coefficients for quartz and clay mineral content prediction reached 0.8106 and 0.8529, respectively, with normalized root mean square error of approximately 0.10. This provides effective technical means for real-time lithology identification and mechanical parameter prediction during conglomerate formation drilling processes. By correlating downhole sensor data with the model’s required mechanical inputs, it becomes possible to perform on-the-fly mineralogy prediction. This capability would enable the dynamic optimization of drilling parameters to adapt to changing formation hardness, providing a powerful technical means to enhance efficiency and safety during drilling in complex conglomerate formations.