1. Bibliographic Information

1.1. Title

The central topic of the paper is "Rock Classification through Knowledge-Enhanced Deep Learning: A Hybrid Mineral-Based Approach".

1.2. Authors

The authors are Iye Szin Anga, Martin Johannes Findl, Elisabeth Hauzinger, Klaus Philipp Sedlazeck, Jyrki Savolainen, Ronald Bakker, Robert Galler, and Elmar Rueckert. Their research backgrounds and affiliations are primarily with Montanuniversität Leoben in Austria, specifically the Chair of Cyber-Physical-Systems, Chair of Subsurface Engineering, and Chair of Resource Mineralogy. One author is also affiliated with LUT-kauppakorkeakoulu in Finland. These affiliations suggest a multidisciplinary team combining expertise in cyber-physical systems, subsurface engineering, and resource mineralogy, which aligns with the paper's focus on integrating deep learning with geological domain knowledge.

1.3. Journal/Conference

The paper is published as a preprint on arXiv, indicated by the provided original source link and publication date. It is not specified whether it has been accepted or published in a peer-reviewed journal or conference at the time of this analysis. arXiv is a well-known open-access repository for preprints of scientific papers, particularly in physics, mathematics, computer science, and related fields. It allows researchers to share their work rapidly before or during the peer-review process, making it highly influential for disseminating new research, though preprints have not yet undergone formal peer review.

1.4. Publication Year

The paper was published on 2025-10-15 (UTC) according to the arXiv publication timestamp.

1.5. Abstract

The abstract outlines the significant challenge of automated rock classification based on mineral composition, highlighting its importance for material recycling, resource management, and industrial processing. It notes that while existing One-dimensional Convolutional Neural Network (1D-CNN) methods are effective for mineral identification via Raman spectroscopy, the complex task of classifying rock types from mineral assemblages remains largely unresolved. This is primarily due to the fact that similar minerals can form different rock types depending on their proportions and formation conditions. To address this, the study introduces a novel knowledge-enhanced deep learning approach that fuses geological domain expertise with spectral analysis. The performance of five machine learning (ML) methods was assessed for mineral classification, with the 1D-CNN and its uncertainty-aware variant demonstrating superior accuracy (98.37±0.006% and 97.75±0.010%, respectively). When the integrated system was evaluated on rock samples, it exhibited varied performance across different lithologies, achieving optimal results for limestone classification but reduced accuracy for rocks composed of similar mineral assemblages. The authors conclude that these findings underscore the inherent challenges in automated geological classification while simultaneously offering a methodological framework to advance material characterization and sorting technologies.

1.6. Original Source Link

The official source for this paper is its arXiv preprint page: https://arxiv.org/abs/2510.13937. The PDF link is: https://arxiv.org/pdf/2510.13937v1.pdf. This clarifies its publication status as a preprint.

2. Executive Summary

2.1. Background & Motivation

The core problem the paper aims to solve is the automated classification of rock types directly from their mineral composition, as identified by Raman spectroscopy. This is a critical challenge with broad implications for industries such as construction, mining, material recycling, resource management, and industrial processing.

The problem is important because traditional rock classification relies heavily on expert geological analysis, involving macroscopic observations, microscopic examinations, and manual mineral identification, often supplemented by chemical data. This process is time-consuming, subjective, and difficult to scale for rapid, automated applications. Automated methods offer the potential for faster, more accurate, and objective classification, which is crucial for sustainable resource management and efficient industrial practices.

Specific challenges and gaps in prior research include:

• While Raman spectroscopy combined with machine learning has achieved high accuracy (over 96%) in identifying individual minerals, there is a fundamental methodological gap in translating these mineral assemblages into rock type classifications.

• The same minerals can occur in different proportions or under different formation conditions to form distinct rock types (e.g., granite vs. sandstone, both containing quartz and feldspar). Existing systems struggle with this nuance.

• A lack of automated systems that can deduce rock types from identified mineral assemblages, which is the "crucial step" that remains unsolved.

  The paper's entry point and innovative idea is to develop a novel knowledge-enhanced deep learning approach. This hybrid method integrates spectral analysis (using 1D-CNN for mineral identification) with geological domain expertise (encoded into a rule-based expert system). By combining data-driven learning with established geological knowledge, the paper seeks to bridge the gap between accurate mineral identification and robust rock type classification, particularly addressing the challenge of varying mineral proportions.

2.2. Main Contributions / Findings

The paper makes several primary contributions:

1. Integrated Framework: It proposes a hybrid integrated framework that combines a One-dimensional Convolutional Neural Network (1D-CNN) for automated mineral identification with a knowledge-enhanced expert system. This system systematically incorporates domain expertise through knowledge integration, leveraging both data-driven learning and established geological knowledge to overcome limitations of traditional rule-based systems.

2. Quantitative Rock Type Classification Methodology: A novel percentage-based mineral composition weighting system is developed for rock type classification. This system accounts for variations in mineral assemblages and their relative abundances, providing a more robust classification framework than traditional binary approaches by considering nuanced compositional differences.

3. Geologically Validated Dataset: The framework is validated using a curated dataset of mineral spectra from the RRUFF database, structured according to expert-designed rock composition templates. This ensures geological validity through systematic sampling of diagnostic mineral assemblages, expert-verified compositional relationships, and high-quality spectral data.

   Key conclusions or findings reached by the paper include:

• For mineral classification, the 1D-CNN achieved a high accuracy of 98.37±0.006%, and its uncertainty-aware variant achieved 97.75±0.010%, significantly outperforming traditional machine learning baselines (SVM, Random Forest, MLP).

• For rock classification using the integrated system, performance varied across lithologies. Limestone classification yielded optimal results with a precision of 66.7%, recall of 57.1%, and an F1-score of 0.62.

• The system showed reduced accuracy for rocks sharing similar mineral assemblages (e.g., granite and sandstone), highlighting a fundamental challenge in differentiating compositionally similar rock types based solely on mineral proportions.

• The knowledge-enhanced approach effectively compensates for data sparsity in rock classification by integrating expert rules.

• The findings emphasize the critical challenges in automated geological classification systems, particularly the mismatch between single mineral spectra and whole rock assemblages, and the difficulty in handling compositional uncertainty.

  These findings provide a methodological framework for advancing material characterization and sorting technologies by demonstrating the feasibility and challenges of automated mineral-to-rock classification.

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

To understand this paper, a novice reader should be familiar with the following fundamental concepts:

• Raman Spectroscopy:

  • Conceptual Definition: Raman spectroscopy is a non-destructive chemical analysis technique that provides detailed information about the chemical structure, phase and polymorphism, crystallinity, and molecular interactions in a sample. It relies on the inelastic scattering of monochromatic light (e.g., laser light) when it interacts with a material. Most of the scattered light is elastically scattered (Rayleigh scattering), meaning its wavelength is unchanged. However, a small fraction of the light undergoes Raman scattering, where its wavelength is shifted due to energy exchange with the vibrational modes of the molecules in the sample.
  • Purpose: The shifts in wavelength (Raman shifts) are unique to the chemical bonds and crystal structure of the material, creating a distinctive "spectral fingerprint" that allows for precise identification of substances, including minerals. In this paper, it is used to identify individual minerals within a rock sample.

• Machine Learning (ML):

  • Conceptual Definition: Machine learning is a subfield of artificial intelligence (AI) that enables computer systems to "learn" from data without being explicitly programmed. Instead of writing code for every possible scenario, ML algorithms build models based on sample data (training data) to make predictions or decisions.
  • Purpose: In this paper, ML is used to classify mineral spectra and, in an integrated system, to classify rock types.

• Deep Learning:

  • Conceptual Definition: Deep learning is a specialized branch of machine learning that uses artificial neural networks with multiple layers (hence "deep") to learn complex patterns from data. These networks are inspired by the structure and function of the human brain.
  • Purpose: Deep learning models, particularly 1D-CNNs, are highly effective for processing sequential data like spectral data due to their ability to automatically learn hierarchical features.

• One-dimensional Convolutional Neural Network (1D-CNN):

  • Conceptual Definition: A Convolutional Neural Network (CNN) is a type of deep learning model commonly used for analyzing visual imagery (2D CNNs). A 1D-CNN, as used in this paper, is adapted for processing sequential or time-series data, such as Raman spectra. It applies convolutional filters (small learnable matrices) across the 1D input data to detect local patterns (e.g., specific peaks or features in a spectrum).
  • Components:
    • Convolutional Layers: Perform convolutions, which involve sliding a filter over the input data and computing a dot product. This extracts features.
    • Activation Functions (e.g., ReLU): Introduce non-linearity into the model, allowing it to learn more complex patterns. ReLU (Rectified Linear Unit) outputs the input directly if it's positive, otherwise it outputs zero.
    • Pooling Layers (e.g., Max Pooling): Reduce the dimensionality of the feature maps, making the model more robust to small shifts in features and reducing computational cost.
    • Fully Connected Layers: Standard neural network layers that take the output from the convolutional and pooling layers and perform the final classification.
  • Purpose: In this paper, 1D-CNNs are specifically used for mineral identification from Raman spectra.

• Uncertainty-Aware Model (1D-CNN-UNK):

  • Conceptual Definition: An uncertainty-aware model is a machine learning model designed not only to make predictions but also to quantify its confidence or uncertainty in those predictions. This is crucial in real-world applications where ambiguous or unknown inputs might occur.
  • Purpose: The 1D-CNN-UNK variant in this paper is specifically designed to handle ambiguous cases and potential unknown mineral assemblages by identifying when it encounters mineral spectra that do not match its predefined classes, essentially signaling "I don't know" or "other."

• Support Vector Machine (SVM):

  • Conceptual Definition: SVM is a powerful supervised learning model used for classification and regression tasks. It works by finding the optimal hyperplane that best separates data points of different classes in a high-dimensional space. The "optimal" hyperplane maximizes the margin between the closest data points of different classes (called support vectors).
  • Purpose: Used as a baseline machine learning model for comparison in mineral classification.

• Random Forest (RF):

  • Conceptual Definition: Random Forest is an ensemble learning method for classification and regression. It operates by constructing a multitude of decision trees during training and outputting the class that is the mode of the classes (classification) or mean prediction (regression) of the individual trees. It reduces overfitting and improves accuracy by averaging the predictions of multiple trees, each trained on a random subset of the data and features.
  • Purpose: Used as a baseline machine learning model for comparison in mineral classification.

• Multilayer Perceptron (MLP):

  • Conceptual Definition: MLP is a foundational type of artificial neural network. It consists of an input layer, one or more hidden layers, and an output layer. Each layer contains multiple neurons, and connections between neurons have associated weights and biases. Information flows forward through the network, and activation functions are applied at each neuron. MLPs are capable of learning non-linear relationships.
  • Purpose: Used as a baseline machine learning model for comparison in mineral classification.

• Evaluation Metrics:

  • Accuracy: The proportion of correctly classified instances out of the total instances.
  • Precision: The ratio of true positive predictions to the total positive predictions (true positives + false positives). It measures the accuracy of positive predictions.
  • Recall (Sensitivity): The ratio of true positive predictions to the total actual positive instances (true positives + false negatives). It measures the ability of the model to find all positive instances.
  • F1-score: The harmonic mean of precision and recall. It provides a single score that balances both precision and recall, which is useful when there's an uneven class distribution.
  • Cross-entropy Loss: A common loss function used in classification tasks with deep learning models. It quantifies the difference between the predicted probability distribution and the true distribution. The goal during training is to minimize this loss.
  • Adam Optimizer: An optimization algorithm used to update the weights and biases of a neural network during training. It combines the advantages of AdaGrad (which works well with sparse gradients) and RMSProp (which works well in non-stationary objectives). It's known for its efficiency and good performance in practice.

3.2. Previous Works

The paper contextualizes its work by reviewing advancements in Raman spectroscopy and machine learning for mineral identification.

• Raman Spectroscopy for Mineral Identification:

  • The application of Raman spectroscopy for rock and mineral identification has seen significant progress, particularly with the advent of AI-driven Raman spectroscopy. This has led to efficient and accurate automated identification of minerals in geological samples, as highlighted by works like Qi et al. [2] and others [5]. This efficiency is crucial for various applications, including resource management.

• Spectral Databases:

  • The development of comprehensive spectral databases is identified as crucial. The RRUFF database [6] is explicitly mentioned as a cornerstone in this domain. It provides quality-controlled data, detailed crystallographic information, and documentation of sample origins.
  • The RRUFF database has been used in diverse applications, from portable gemstone identification systems [7] to various machine learning and deep learning approaches [8, 9]. Recent developments have further expanded its utility through high-throughput computational methods [10] and open-source analysis tools [11]. The current paper utilizes the RRUFF database for its training data.

• Machine Learning for Raman Spectrum Analysis:

  • Qi et al. [2] provide a review of machine learning advances in Raman spectrum data analysis, covering traditional statistical methods to deep learning.
  • 1D-CNNs are specifically noted as highly effective architectures for spectral data analysis [1], demonstrating excellent performance across various spectroscopic applications [12, 13]. This forms the basis for the mineral classification component of the proposed system.

• Industrial Applications and Sustainability:

  • The practical application of automated Raman analysis extends to industrial settings, particularly in sustainable resource management. Res et al. [4] demonstrate its utility in characterizing excavated materials, showing that precise characterization and classification can enable the recycling of secondary raw materials, contributing to sustainable construction practices.

3.3. Technological Evolution

The field has evolved from traditional manual, expert-driven geological classification to automated spectral analysis aided by machine learning.

1. Manual/Expert-Based Classification: Historically, rock classification relied on geologists' expertise, involving visual inspection, microscopy, and chemical tests. This is accurate but slow, subjective, and not scalable.

2. Spectroscopic Techniques: The introduction of techniques like Raman spectroscopy provided objective, non-destructive, and rapid means to identify minerals based on their unique spectral fingerprints.

3. Early Machine Learning: Initial integration involved traditional machine learning algorithms (SVM, Random Forest, MLP) to analyze Raman spectra for mineral identification. These methods improved automation but might struggle with complex, high-dimensional spectral data.

4. Deep Learning for Spectroscopy: The advent of deep learning, particularly 1D-CNNs, revolutionized spectral data analysis. These models can automatically learn intricate features from raw spectra, leading to highly accurate mineral identification.

5. Current Gap (Motivation of this paper): Despite highly accurate mineral identification, the critical step of deducing rock types from these identified mineral assemblages remained largely unsolved. This is because rock types are defined not just by the presence of minerals, but by their proportions and geological context.

   This paper's work fits into the technological timeline as the next logical step: bridging the gap between high-accuracy mineral identification (achieved by 1D-CNNs) and robust rock type classification by integrating geological domain knowledge (expert rules) into the deep learning pipeline.

3.4. Differentiation Analysis

Compared to the main methods in related work, this paper's approach offers several core differences and innovations:

• Integration of Knowledge and Data-Driven Learning: Prior research primarily focused on either mineral-level identification using Raman spectroscopy and machine learning or traditional expert systems based solely on predefined rules. This paper's core innovation is a hybrid approach that explicitly combines the strengths of 1D-CNNs (for high-accuracy mineral spectral analysis) with a knowledge-enhanced expert system (for incorporating geological domain expertise regarding mineral assemblages and proportions). This addresses the limitations of both purely data-driven models (which might lack geological context) and purely rule-based systems (which might struggle with spectral interpretation).
• Addressing the Mineral-to-Rock Deduction Gap: The paper directly tackles the previously unsolved challenge of automatically classifying rock types from identified mineral assemblages. This is distinct from mineral identification, as it requires understanding how combinations and proportions of minerals define a rock type, which is a key differentiator from existing mineral identification systems.
• Quantitative Compositional Weighting System: The development of a percentage-based mineral composition weighting system is novel. This moves beyond simple presence/absence or binary classification by quantitatively considering the relative abundances of key minerals, allowing for a more nuanced and geologically robust classification, especially for rocks with similar mineral constituents but different proportions.
• Uncertainty Handling: The inclusion of an uncertainty-aware 1D-CNN variant and confidence thresholds in the expert system explicitly acknowledges and attempts to manage ambiguity and uncertainty in geological classification, which is a practical necessity often overlooked in simpler models. This allows the system to identify cases where classification confidence is low or where a sample might represent an "unknown" type.
• Geologically-Informed Dataset Curation: Instead of relying solely on generic data augmentation, the paper emphasizes a geologically-informed sampling strategy for its dataset, ensuring representativeness and relevance to real-world geological conditions. This thoughtful approach to data preparation is crucial for developing robust geological classification systems.

4. Methodology

4.1. Principles

The core idea of the method is to create a robust rock classification framework by integrating the strengths of data-driven deep learning (specifically 1D-CNN for mineral identification) with geological domain expertise encoded in a rule-based expert system. The theoretical basis is that while Raman spectroscopy can accurately identify individual minerals, rock type classification requires understanding the mineral assemblages and their relative proportions, which is best captured by established geological knowledge.

The intuition behind this hybrid approach is to first leverage the power of deep learning to accurately identify individual minerals from their Raman spectra. Then, instead of solely relying on the deep learning model to deduce the rock type (which might struggle with the contextual nuances of mineral proportions that define rock types), the identified minerals are fed into an expert system. This expert system embodies the accumulated knowledge of geologists regarding what combinations and proportions of minerals constitute specific rock types. By combining these two layers, the system aims to overcome the limitations of each approach when used in isolation, leading to more accurate and geologically sound rock classification. An uncertainty-aware component is also integrated to handle ambiguous or unknown cases, making the system more robust for real-world applications.

The following figure (Figure 1 from the original paper) shows the schematic overview of the knowledge-enhanced rock classification system.

该图像是一个示意图，展示了通过知识增强深度学习进行岩石分类的流程。图中包含了岩石样本的多个点测量，矿物检测环节采用了1D-CNN模型和不确定性-aware模型，结合矿物关联规则和置信度评分进行知识整合，最终实现基于置信度的岩石分类。

The workflow involves processing multiple measurement points ( $\mathbf { n } \geq \mathbf { 10 }$ ) from each rock sample. These measurements are fed into either a standard 1D-CNN or an uncertainty-aware variant for mineral detection. The output from mineral detection is then passed to a knowledge-guided system that uses expert-defined association rules and confidence scoring (parameterized by thresholds $\delta \mathrm { c }$ and $\delta \mathrm { d }$). Finally, rock classification is performed based on these confidence metrics. This integrated process combines data-driven learning with domain expertise to address the fundamental challenges in automated rock type classification from Raman spectroscopy measurements.

4.2. Core Methodology In-depth (Layer by Layer)

The methodology encompasses four key components:

1. Classification Framework: Establishes the theoretical foundation for mineral-based rock type identification.
2. Systematic Data Collection and Generation: Procedures for obtaining and expanding the dataset.
3. Development and Implementation of Two Statistical Classifiers: This refers to the 1D-CNN and its uncertainty-aware variant for mineral classification.
4. Knowledge-Guided System: A rule-based expert system that utilizes established geological knowledge for final classification decisions.

4.2.1. System Assumptions and Theoretical Foundations

The hybrid rock classification system operates under several key assumptions:

• The system can effectively integrate Raman spectral data and expert system rules.

• Each Raman measurement point sufficiently captures representative mineral assemblages of the rock sample.

• The presence and relative abundances of key mineral assemblages are sufficient for preliminary rock type determination.

• Expert geological knowledge can be effectively encoded into a rule-based system, including standardized classification schemes and expert experience.

• Natural variations in mineral compositions and assemblages can be accommodated within defined confidence intervals.

  The classification rules in this system are derived from geological literature.

• Igneous Rocks (Plutonic/Volcanic): The classification framework is based on the Quartz-Alkali Feldspar-Plagioclase Feldspar (QAPF) diagrams established by the International Union of Geological Sciences (IUGS) [15, 16]. In this scheme, modal mineral contents (the volume percentages of minerals in a rock) are normalized to 100% and plotted in the QAPF diagram, which places the rock into a corresponding field (e.g., the granite field). This standardized scheme provides the theoretical foundation for the expert system's decision rules, particularly for granite and other plutonic igneous rock classifications.

• Sedimentary Rocks (Sandstone and Limestone): Classification rules for sandstone and limestone were developed through knowledge engineering (the process of acquiring, representing, and using knowledge in an expert system) with expert geologists.

  • Sandstone: The classification for clastic sediments (rocks composed of fragments of pre-existing rocks) with grain sizes < 2 mm typically uses ternary classification diagrams that plot the relative proportions of quartz, feldspar, and lithic fragments (rock fragments), normalized to 100%. The paper specifically mentions schemes developed by Krynine (1948) [17], Dott (1964) [18], and Pettijohn et al. (1972) [19], which are widely accepted (e.g., by Blatt & Frie, 1980 [20]). Additionally, variations in matrix content (fine-grained material between larger grains, typically with grain sizes < 30 µm) are considered along an axis perpendicular to the ternary diagram. Rocks with < 15% matrix are arenites, ≥ 15% matrix are wackes, and ≥ 75% matrix are claystone. The study focuses on quartzitic/quartz sandstones with 0-15% matrix content. The sandstone decision is further refined based on the calcite-dolomite ratio.

  • Limestone: The classification focuses on typical limestone (≤ 10% dolomite) and dolomitic limestone (10-50% dolomite) [20, 23]. This process incorporated standard sedimentary rock classification schemes, expert field identification practices, and practical experience in distinguishing key mineral assemblages and textural/compositional indicators.

    For the paper's mineral-based classification approach, specific rules are:

• Sandstone: Uses quartz, feldspar, and mica compounds, with calcite, pyrite, rutile, and tourmaline as accessory minerals.

• Limestone: Incorporates carbonate mineral assemblages (calcite, dolomite) and common impurities (quartz, feldspar, pyrite).

• Granite: Considers mica compounds in addition to the main minerals (feldspars, quartz).

4.2.2. Rock Classification Framework

The framework addresses the computational challenges of classifying rocks from spectral data by using a dual-layer classification architecture. It covers three rock types: granite (igneous), sandstone (clastic sedimentary), and limestone (chemical/biochemical sedimentary), chosen for their different geological origins and distinct mineral compositions.

4.2.2.1. Classification Overview

• Granite: (Figure 2 shows the hierarchical relationship between Granite and its essential minerals)

  该图像是示意图，展示了花岗岩及其主要矿物的层次关系。图中明确指明了花岗岩的成分比例，包括长石、石英和云母的含量分布。

  Granite is an igneous rock formed from the cooling of magma. Its major rock-forming minerals include feldspars (e.g., Albite, Anorthite, Orthoclase), mica group minerals (e.g., Biotite, Muscovite, Phlogopite), and Quartz. The typical proportions are feldspars (45-80%), quartz (20-40%), and mica minerals (0-15%) [15].

• Sandstone: (Figure 3 shows the hierarchical relationship between Sandstone and its essential minerals)

  该图像是一个示意图，展示了砂岩及其主要矿物的层级关系。图中显示砂岩的组成成分及其相对百分比，包括石灰石、石英、长石、黄铁矿等，帮助理解不同矿物在砂岩中的作用。

  Sandstone is a clastic sedimentary rock predominantly composed of quartz (>70%). It also includes significant contributions from feldspars (5-25%, comprising both alkali feldspars and plagioclase) and minor components such as calcite (<10%), pyrite (<1%), mica group minerals (2-3%), rutile (<2%), and tourmaline (<2%). Calcite, pyrite, rutile, and tourmaline are considered accessory minerals in sandstone.

• Limestone: (Figure 4 shows the hierarchical relationship between Limestone and its essential minerals)

  该图像是示意图，展示了石灰岩与其基本矿物之间的层级关系。图中包括纯石灰岩的成分和多种碳酸盐及沉积矿物的分类，反映了不同矿物在石灰岩分类中的比例要求。

  Limestone is primarily a carbonate rock formed through chemical or biochemical processes, usually in shallow marine environments. It is characterized by a predominant calcite content of >50% and a dolomite content ranging from 0-50% [23]. Minor constituents, such as quartz (<10%), feldspar (<5%, divided into alkali feldspar and plagioclase), and pyrite (<5%), can also be present.

4.2.2.2. Confidence-Based Classification

To manage uncertainty in real-world classification, a confidence-based classification mechanism is introduced. This extends the expert system by incorporating weighted confidence scores and dual thresholds for more robust decision-making. Statistical methods are applied to measure confidence intervals and establish confidence ($\theta_c$) and dominance ($\theta_d$) thresholds (0.7 and 0.3, respectively). These thresholds systematically manage classification uncertainty and balance preliminary mineral type identification with the inherent variability of mineral compositions in geological samples.

The classification decision for a rock type is formalized as: $$C _ { r o c k } = \left\{ \begin{array} { l l } { \mathrm { r o c k \ t y p e, } } & { \mathrm { i f } w _ { m a x } \geq \theta _ { c } \mathrm { a n d } ( w _ { m a x } - w _ { 2 n d } ) \geq \theta _ { d } } \\ { \mathrm { o t h e r, } } & { \mathrm { o t h e r w i s e } } \end{array} \right.$$ Where:

• w _ { m a x } is the highest weight among all rock types.

• w _ { 2 n d } is the second highest weight among all rock types.

• $\theta _ { c }$ is the confidence threshold (set to 0.7).

• $\theta _ { d }$ is the dominance threshold (set to 0.3).

• If the conditions are met, the sample is classified as the rock type corresponding to $w_{max}$; otherwise, it is classified as "other," indicating insufficient certainty or ambiguity.

  This dual-threshold approach prevents misclassification when the highest confidence score is below $\theta_c$ (insufficient certainty) or when the difference between the highest and second-highest weights is less than $\theta_d$ (suggesting ambiguity between rock types). The thresholds are empirically determined through validation to balance classification accuracy with reliability.

The weight calculation for each rock type considers the relative proportions of key minerals, as summarized in Table 1. The following are the results from Table 1 of the original paper: Table 1 Mineral Proportions for Rock Types

To quantify the weight for each rock type based on a sequence of measurements, the following formula is used: For a sequence of measurements $M = \{ m _ { 1 } , . . . , m _ { n } \}$ , the weight for each rock type R _ { l } is calculated as: $$w _ { R _ { l } } = \sum _ { i = 1 } ^ { k } \alpha _ { i } \cdot \delta _ { i }$$ Where:

• w _ { R _ { l } } is the calculated weight for rock type $R_l$.

• $k$ is the total number of relevant minerals considered for the rock type $R_l$.

• $\alpha _ { i }$ is the $i$-th mineral's abundance (or weighting coefficient representing relative importance) in determining the rock type classification.

• $\delta _ { i }$ is an indicator function that equals 1 if the proportion of the $i$-th mineral in the measurements falls within its expected minimum and maximum range ($p _ { i } ^ { m i n } \leq f _ { i } ( M ) \leq p _ { i } ^ { m a x }$), and 0 otherwise.

• f _ { i } ( M ) is the observed proportion of the $i$-th mineral in the sequence of measurements $M$.

• $p _ { i } ^ { m i n }$ and $p _ { i } ^ { m a x }$ represent the minimum and maximum expected proportions for the $i$-th mineral for a specific rock type $R_l$.

  This formula ensures that the classification proportionally accounts for the presence and relative abundance of minerals, providing a robust framework for automated rock type identification.

4.2.3. Dataset Collection and Generation

The dataset primarily originates from the RRUFF database, which contains approximately 7,000 mineral samples representing 3,500 distinct mineral species. The authors note a significant class imbalance within the database, with 1,522 mineral classes having limited samples.

Rather than standard data augmentation, a geologically-informed sampling strategy was adopted. This strategy involved selecting specific samples from the RRUFF database based on geological context, ensuring the dataset is relevant to real-world geological conditions. This approach also complements the hybrid architecture by integrating expert knowledge to compensate for limited training samples. The chosen spectra for analysis are from conditions where single crystals were typically used, resulting in typical random crystallographic orientations of the present mineral phases.

Due to the limited sample size from this targeted selection, the dataset was expanded using two synthetic data generation methods:

1. PCA-based approach: Applied for minerals with larger existing datasets. Principal Component Analysis (PCA) is a statistical procedure that uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of linearly uncorrelated variables called principal components. This can be used to generate synthetic data by sampling along the principal components.

2. Direct variation method: Used for minerals with limited existing samples. This likely involves simpler perturbation or interpolation techniques to create new samples based on existing ones.

   The target dataset sizes were determined by applying a $4 \times$ multiplication factor to the original sample number, with specific projections for feldspar, quartz, mica, calcite, and other minerals relevant to granite, sandstone, and limestone. Standard preprocessing steps, such as background subtraction and peak fitting, were also applied, though specific methods might vary as RRUFF aggregates data from multiple institutions.

The effectiveness of this geological context-driven selection approach was validated through expert-designed test cases (detailed in Section 3.5.5 of the paper).

4.2.4. Mineral Classification

For mineral classification, five machine learning models were implemented and tested using an expanded dataset of 1366 mineral samples:

• Support Vector Machine (SVM)

• Random Forest (RF)

• Multilayer Perceptron (MLP)

• One-dimensional Convolutional Neural Network (1D-CNN)

• Uncertainty-aware 1D-CNN (1D-CNN-UNK)

  Two versions of the 1D-CNN architecture were developed:

• Base 1D-CNN model: Consists of two convolutional layers (with 16 and 32 channels, respectively) using ReLU activation and max pooling operations. The network processes input spectra to analyze their features. The output layer uses a softmax activation function for multiclass classification, assigning probabilities to each of the predefined mineral classes.

• Uncertainty-aware 1D-CNN (1D-CNN-UNK): This variant is designed to identify when the model encounters mineral spectra that do not match the predefined mineral classes. This is achieved by adding an "unknown" class to the model's output, enabling it to flag ambiguous or novel spectra.

  Both 1D-CNN models were trained using cross-entropy loss and the Adam optimizer with a learning rate of 0.01. Early stopping was implemented with a patience of 20 epochs, meaning training was halted if the validation loss did not improve for 20 consecutive epochs, preventing overfitting.

4.2.5. Rule-Based Expert System

The expert system was developed to automate rock classification based on Raman spectroscopy point measurements by leveraging expert knowledge effectively. It applies a hierarchical decision-making process that considers both prediction accuracy and the presence of mineral assemblages characteristic of different rock types.

4.2.5.1. Knowledge Base and Definitions

The knowledge base KB is formally defined as a quintuple: $KB = ( G , R , H , P , C )$ Where:

• $G = \{ G _ { 1 } , G _ { 2 } , . . . , G _ { K } \}$ represents $K$ distinct mineral assemblages (groups of minerals that commonly occur together).

• $R = \{ R _ { 1 } , R _ { 2 } , . . . , R _ { L } \}$ denotes $L$ rock type classification rules. Each rule defines criteria for a specific rock type.

• $H = ( V , E )$ defines the hierarchical decision tree structure (e.g., how mineral groups relate to rock types, or how larger mineral categories break down into specific minerals). $V$ represents vertices (nodes) and $E$ represents edges (connections) in the tree.

• $P = \{ p _ { v } | v \in V \}$ comprises classification and composition parameters associated with each node in the decision tree.

• $C = \{ C _ { 1 } , C _ { 2 } , . . . , C _ { J } \}$ represents $J$ compositional and confidence constraints (e.g., minimum/maximum percentages of certain minerals, or overall confidence thresholds).

  For any mineral $m$ and group G _ { i } with weight w _ { i } , the weighted membership function is defined as: $$\mu _ { G _ { i } } ( m , w _ { i } ) = \left\{ \begin{array} { l l } { w _ { i } } & { \mathrm { i f } m \in G _ { i } \mathrm { a n d } p _ { i } ^ { m i n } \leq f _ { i } ( m ) \leq p _ { i } ^ { m a x } } \\ { 0 } & { \mathrm { o t h e r w i s e } } \end{array} \right.$$ Where:

• $\mu _ { G _ { i } } ( m , w _ { i } )$ is the membership score of mineral $m$ to mineral group $G_i$, weighted by $w_i$.

• w _ { i } is the weight assigned to the mineral group $G_i$ in a particular context (e.g., its importance for a rock type).

• $m \in G _ { i }$ checks if the mineral $m$ belongs to the mineral group $G_i$.

• f _ { i } ( m ) is the observed proportion of mineral $m$ (or minerals from group $G_i$) in the measurements.

• $p _ { i } ^ { m i n }$ and $p _ { i } ^ { m a x }$ are the minimum and maximum expected proportions for minerals from group $G_i$ for a given rock type.

• The function returns $w_i$ only if the mineral belongs to the group AND its proportion is within the defined range; otherwise, it returns 0.

4.2.5.2. Compositional Rules and Constraints

The confidence-based compositional requirements for each rock type are formalized through constraint functions: $$C _ { j } ( M , G _ { i } , w _ { R _ { l } } ) = f _ { j } ( n _ { i } , \alpha _ { j } , w _ { R _ { l } } ) \ge \beta _ { j }$$ Where:

• C _ { j } ( M , G _ { i } , w _ { R _ { l } } ) is the $j$-th constraint function for a rock type rule $R_l$.

• $M$ represents the sequence of measurements.

• G _ { i } is a mineral group.

• w _ { R _ { l } } is the importance weight associated with the rock type classification rule $R_l$.

• f _ { j } represents a constraint function (e.g., a function calculating a score based on mineral counts and weights).

• n _ { i } is the count of minerals from group G _ { i } found in the measurements $M$.

• $\alpha _ { j }$ is a parameter vector associated with the $j$-th constraint (e.g., containing coefficients for calculating the constraint value).

• $\beta _ { j }$ defines the threshold value that the constraint function must meet or exceed.

  The classification rule R _ { l } (for a specific rock type) incorporating confidence thresholds is expressed as: $$R _ { l } ( M ) = \left( \bigwedge _ { j \in \mathcal { I } _ { l } } C _ { j } ( M , G _ { i _ { j } } , w _ { i _ { j } } ) \right) \wedge C _ { c o n f } ( w _ { m a x } , w _ { 2 n d } )$$ Where:

• R _ { l } ( M ) is the logical evaluation of rule $R_l$ given measurements $M$.

• $\bigwedge$ is the logical AND operator over a sequence of conditions.

• $j \in \mathcal { I } _ { l }$ indicates that the rule $R_l$ is satisfied if all individual compositional constraints C _ { j } relevant to $R_l$ are met.

• C _ { j } ( M , G _ { i _ { j } } , w _ { i _ { j } } ) is the $j$-th compositional constraint related to mineral group $G_{i_j}$ and its importance weight $w_{i_j}$.

• C _ { c o n f } ( w _ { m a x } , w _ { 2 n d } ) represents the confidence constraint (as defined by the $\theta_c$ and $\theta_d$ thresholds for $w_{max}$ and $w_{2nd}$). This ensures that the overall classification also meets the required confidence and dominance criteria.

4.2.5.3. Mineral Assemblages

Mineral assemblages for each rock type R _ { l } are represented as weighted sets with composition ranges: $$A _ { R _ { l } } = \{ ( G _ { i } , w _ { i } , [ p _ { i } ^ { m i n } , p _ { i } ^ { m a x } ] ) | G _ { i } \in G , w _ { i } \in [ 0 , 1 ] , p _ { i } ^ { m i n } , p _ { i } ^ { m a x } \in [ 0 , 1 ] \}$$ Where:

• A _ { R _ { l } } is the mineral assemblage definition for rock type $R_l$.
• $G _ { i } \in G$ indicates that $G_i$ is a mineral group from the set of all mineral assemblages.
• $w _ { i } \in [ 0 , 1 ]$ is the weight representing the importance of mineral group $G_i$ for classifying rock type $R_l$.
• $[ p _ { i } ^ { m i n } , p _ { i } ^ { m a x } ] \in [ 0 , 1 ]$ is the expected compositional range (minimum and maximum proportions) for mineral group $G_i$ within rock type $R_l$.

4.2.5.4. Weighted Importance Model

Given the weighted mineral assemblages A _ { R _ { l } }, the probability of observing rock type R _ { l } given measurements $M$ is calculated as: $$P ( R _ { l } | M ) = \prod _ { ( G _ { i } , w _ { i } , [ p _ { i } ^ { m i n } , p _ { i } ^ { m a x } ] ) \in A _ { R _ { l } } } w _ { i } ^ { n _ { i } } \cdot \delta _ { i }$$ Where:

• $P ( R _ { l } | M )$ is the likelihood or score for rock type $R_l$ given the measurements $M$.

• $\prod$ is the product operator over all mineral groups $G_i$ that are part of the assemblage $A_{R_l}$.

• w _ { i } is the importance weight for mineral group $G_i$, derived from geological composition ranges.

• $n _ { i } = \mathrm { count } ( M , G _ { i } )$ is the number of individual mineral identifications from group G _ { i } in the measurements $M$.

• $\delta _ { i }$ is an indicator function that equals 1 if the observed proportion of mineral group $G_i$ in $M$ (f _ { i } ( M )) is within its defined range ($p _ { i } ^ { m i n } \leq f _ { i } ( M ) \leq p _ { i } ^ { m a x }$), and 0 otherwise. This ensures that only mineral groups contributing within their expected ranges positively influence the score.

  This model calculates a score for each rock type based on the multiplicative effect of the importance weights of the identified minerals, modulated by whether their proportions fall within the expert-defined geological ranges.

4.2.5.5. Evaluation Framework

To evaluate the expert system's performance, 10 test cases for each rock type (granite, sandstone, limestone) were utilized. These cases were specifically designed by an expert geologist to represent both confident and non-confident classification scenarios.

• Confident Cases: Involved mineral assemblages that unambiguously indicate specific rock types, representing typical compositions found in well-documented geological formations.
• Non-Confident Cases: Designed with mineral assemblages that were ambiguous, borderline, or lacked strong indicators for any specific rock type, or specifically designed to reject a type. The evaluation used standard metrics including accuracy, precision, recall, and F1-score to provide a comprehensive assessment of the system's classification capabilities.

5. Experimental Setup

5.1. Datasets

The experiments primarily utilized the RRUFF database (https://rruff.info/zipped_data_files/) as the source for mineral spectra.

• Source and Characteristics:

  • RRUFF database: A comprehensive, quality-controlled repository of Raman spectra and associated crystallographic data for minerals. It contains approximately 7,000 mineral samples representing 3,500 distinct mineral species. The paper used high-quality unoriented Raman spectra from RRUFF.
  • Class Imbalance: The authors identified a significant class imbalance within RRUFF, with 1,522 mineral classes having limited samples.
  • Geologically-informed Sampling: Instead of generic data augmentation, a geologically-informed sampling strategy was employed. This involved selecting specific spectra samples from RRUFF that are relevant to real-world geological conditions and align with the hybrid architecture's expert knowledge integration. The selected spectra were from conditions where single crystals were used, resulting in typical random crystallographic orientations.

• Scale and Domain:

  • Expanded Dataset for Mineral Classification: For training the machine learning models (SVM, RF, MLP, 1D-CNN, 1D-CNN-UNK), the dataset was expanded to 1366 mineral samples. This expansion was achieved using two synthetic data generation methods: a PCA-based approach for minerals with larger datasets and a direct variation method for minerals with limited samples. The target dataset sizes were a $4 \times$ multiplication factor of the original samples, specifically for minerals relevant to granite, sandstone, and limestone (i.e., feldspars, quartz, mica, calcite).
  • Test Cases for Rock Classification: For evaluating the integrated system's rock classification performance, a limited dataset of 30 rock samples (10 for each rock type: granite, sandstone, limestone) was used. These test cases were expert-designed to represent both confident and non-confident mineral assemblages, ensuring geological validity.

• Effectiveness for Validation:

  • The RRUFF database is highly effective for validating mineral identification due to its standardized, high-quality spectral data.
  • The expert-designed rock composition templates and specific test cases (confident and non-confident) for rock classification are crucial for validating the knowledge-enhanced approach's ability to interpret mineral assemblages within a geological context, especially given the data sparsity acknowledged by the authors. The choice of granite, sandstone, and limestone allows for testing across igneous and sedimentary rock types with varying mineralogical complexities.

5.2. Evaluation Metrics

The paper uses several standard classification metrics to evaluate the performance of both the mineral classification models and the integrated rock classification system.

For all metrics, we define:

• TP (True Positives): Instances correctly predicted as positive.
• TN (True Negatives): Instances correctly predicted as negative.
• FP (False Positives): Instances incorrectly predicted as positive.
• FN (False Negatives): Instances incorrectly predicted as negative.

1. Accuracy

   • Conceptual Definition: Accuracy measures the overall correctness of the model's predictions. It is the proportion of the total number of predictions that were correct.
   • Mathematical Formula: $ \text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} $
   • Symbol Explanation:
     • TP: The number of instances correctly identified as positive.
     • TN: The number of instances correctly identified as negative.
     • FP: The number of instances incorrectly identified as positive (Type I error).
     • FN: The number of instances incorrectly identified as negative (Type II error).

2. Precision

   • Conceptual Definition: Precision measures the proportion of positive identifications that were actually correct. It answers the question: "Of all the instances the model predicted as positive, how many were truly positive?" High precision indicates a low rate of false positives.
   • Mathematical Formula: $ \text{Precision} = \frac{TP}{TP + FP} $
   • Symbol Explanation:
     • TP: The number of instances correctly identified as positive.
     • FP: The number of instances incorrectly identified as positive.

3. Recall (Sensitivity)

   • Conceptual Definition: Recall measures the proportion of actual positives that were correctly identified. It answers the question: "Of all the instances that were actually positive, how many did the model correctly identify?" High recall indicates a low rate of false negatives.
   • Mathematical Formula: $ \text{Recall} = \frac{TP}{TP + FN} $
   • Symbol Explanation:
     • TP: The number of instances correctly identified as positive.
     • FN: The number of instances incorrectly identified as negative.

4. F1-score

   • Conceptual Definition: The F1-score is the harmonic mean of precision and recall. It provides a single score that balances both precision and recall, which is particularly useful when there is an uneven class distribution (i.e., a large number of actual negatives). An F1-score of 1 is perfect precision and recall, while 0 means either one or both are zero.
   • Mathematical Formula: $ \text{F1-score} = 2 \cdot \frac{\text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}} $
   • Symbol Explanation:
     • $\text{Precision}$: The precision calculated for the model.
     • $\text{Recall}$: The recall calculated for the model.

5.3. Baselines

For mineral classification, the paper compared its 1D-CNN models against three traditional machine learning baselines:

• Support Vector Machine (SVM)
• Random Forest (RF)
• Multilayer Perceptron (MLP) These baselines are representative as they are widely used and well-established algorithms for classification tasks, providing a solid reference point to demonstrate the efficacy of the deep learning approach.

For the integrated system performance in rock classification, the comparison was between:

• A baseline model (implicitly the knowledge-guided 1D-CNN without explicit uncertainty handling, or the standard 1D-CNN integrated into the expert system without the "unknown" class).
• The uncertainty-aware variant (uncertainty-aware knowledge-guided 1D-CNN or 1D-CNN-UNK integrated into the expert system). This comparison helps to specifically evaluate the impact of uncertainty handling on the overall rock classification performance.

6. Results & Analysis

The experimental evaluation of the mineral assemblage-based classification framework covers mineral classification and the performance of the integrated system.

6.1. Core Results Analysis

6.1.1. Mineral Classification

The performance of mineral classification was evaluated using the mean accuracy across five-fold cross-validation.

The following are the results from Figure 5 of the original paper:

该图像是一个柱状图，展示了五种机器学习模型的分类准确率。结果显示，1D-CNN 和其不确定性感知变体的准确率最高，分别为 98.37% 和 97.75%。其他模型的准确率依次是 SVM 0.8843，Random Forest 0.9585，MLP 0.9625。

As shown in the bar chart (Figure 5), the 1D-CNN model achieved the highest accuracy at 98.37%, while the 1D-CNN-UNK model showed a slightly lower, but still excellent, performance at 97.75%. Both deep learning models significantly outperformed the traditional machine learning baseline approaches:

• Support Vector Machine (SVM): 88.43%

• Random Forest (RF): 95.85%

• Multilayer Perceptron (MLP): 96.25%

  This strongly validates the effectiveness of the 1D-CNN architecture for mineral identification from Raman spectra. While a confusion matrix would provide more detailed per-mineral performance, the high overall accuracy of the 1D-CNN models suggests their efficacy as the initial mineral detection layer in the integrated system.

6.1.2. Integrated System Performance

The integrated system's performance was evaluated using confusion matrices (Figure 6) for both the baseline knowledge-guided 1D-CNN and the uncertainty-aware knowledge-guided 1D-CNN over 30 rock samples ($\Delta \mathrm{n} = 30$).

The following are the results from Figure 6 of the original paper:

该图像是混淆矩阵的示意图，展示了知识引导1D-CNN（左）和不确定性感知知识引导1D-CNN（右）在岩石分类中的性能。矩阵中的每个单元显示了真实标签与预测标签之间的对应关系，有助于分析模型的分类能力。

The comparative analysis reveals distinct performance patterns across rock classifications.

Granite Classification:

• Baseline Model (Figure 6a): Achieves 33.3% precision and 100% recall for granite identification (n=5 samples). A 100% recall means all actual granite samples were identified, but low precision indicates many non-granite samples were incorrectly classified as granite.
• Uncertainty-Aware Variant (Figure 6b): Exhibits 23.5% precision and 80% recall. The precision is even lower, and recall slightly reduced, suggesting the uncertainty-aware model is more conservative or flags more ambiguous cases as "other" or misclassifies more often for granite.

Limestone Classification:

• Both Models: Maintain consistent performance metrics for limestone samples (n=7):
  • Precision: 66.7%
  • Recall: 57.1%
  • F1-score: 0.62 This indicates that limestone is relatively well-classified by both variants, likely due to its distinct mineral assemblage dominated by calcite.

Sandstone Classification:

• Baseline Model (Figure 6a): Achieves 57.1% precision, 36.4% recall, and an F1-score of 0.44 for sandstone samples (n=11).
• Uncertainty-Aware Variant (Figure 6b): Demonstrates reduced classification accuracy with 40% precision, 18.2% recall, and an F1-score of 0.25. The performance degradation for sandstone in the uncertainty-aware variant is notable, particularly the significant drop in recall.

Misclassification Patterns:

• Quantitative analysis of misclassification patterns reveals systematic variations. The uncertainty-aware variant shows increased granite misclassifications, especially for samples from the "Other" category ($\Delta \mathrm{n}=6$ vs. $\mathrm{n}=4$ in baseline) and sandstone samples ($\Delta \mathrm{n}=5$ vs. $\mathrm{n}=4$ in baseline). This pattern suggests that uncertainty-aware decision-making can sometimes lead to more cautious (or perhaps more false negative) classifications for certain types.
• The results highlight two fundamental challenges:

  1. The inherent mismatch between data composition (single mineral spectra) and the target classification objective (whole rock assemblages). This is exemplified by the $\Delta \mathrm{n}=+1$ sample for granite misclassification from "Other" category by the uncertainty-aware model, indicating it might be more prone to labeling complex cases as "Other" or misclassifying when unsure.

  2. While the 1D-CNN architecture effectively learns individual mineral spectral signatures, the classification accuracy decreases when processing complex mineral assemblages within whole rock samples. This is particularly evident in the $\Delta F1_{sandstone} = -0.19$ (a decrease of 0.19 in F1-score for sandstone) for the uncertainty-aware model compared to the baseline. Both models showed limited effectiveness in differentiating between compositionally similar rock types (granite and sandstone), with P(granite) (likely referring to the confidence score or probability of being granite) remaining below 35% for both.

     The observed performance patterns are most significant where rocks share similar mineral constituents in varying proportions, such as granite and sandstone. The uncertainty-aware variant's performance indicates that its primary challenge extends beyond classifying individual minerals to making discerning whole-rock classifications. This points to a need for more comprehensive mineral assemblage training datasets that capture spectral interactions within whole rock samples. Additionally, measuring the relative amounts of different minerals could further enhance the analysis and improve accuracy.

6.2. Limitation

The paper explicitly acknowledges a key limitation: although the method effectively detects the presence of specific minerals through their characteristic Raman spectra, it faces a significant challenge in scaling and accurately classifying rock types, especially when these types share similar mineral assemblages but have distinct geological origins. This is evidenced by the low precision rates for granite classification (<35%) and the systematic misclassification patterns observed between compositionally similar rock types. The small sample size used for rock classification (n=30) also inherently limits the generalizability of these specific rock classification results.

7. Conclusion & Reflections

7.1. Conclusion Summary

This investigation significantly advances automated geological classification by integrating Raman spectroscopy with knowledge-enhanced deep learning. The quantitative analysis, based on a limited dataset of 30 samples for rock classification, demonstrates the methodological feasibility and effectiveness of the proposed hybrid mineral-to-rock classification framework. The 1D-CNN architecture achieved excellent mineral identification accuracy (98.37±0.006%), with its uncertainty-aware variant closely following (97.75±0.010%). The implementation of confidence thresholds within the knowledge system proved crucial for differentiating rock types, particularly limestone, which showed optimal performance with 66.7% precision, 57.1% recall, and an F1-score of 0.62. The framework successfully addresses fundamental challenges by systematically combining spectroscopic data analysis with expert geological knowledge, demonstrating the approach's ability to compensate for data sparsity through rule integration.

7.2. Limitations & Future Work

The authors themselves highlighted several limitations and suggest future research directions:

• Small Sample Size: The most prominent limitation is the small sample size (n=30) used for evaluating rock classification. This significantly limits the generalizability of the findings, and further validation with a larger dataset is explicitly stated as necessary.
• Differentiation of Similar Assemblages: The method struggles with scaling and accurately classifying rock types that share similar mineral assemblages but have distinct geological origins or proportions. This was particularly evident in the low precision rates for granite and the misclassification patterns between compositionally similar rock types like granite and sandstone.
• Data Representation Mismatch: A fundamental challenge identified is the inherent mismatch between data composition (single mineral spectra) and the target classification objective (whole rock assemblages).
• Future Development Pathways:
  • Transition to Industrial Operations: Advancing from controlled laboratory conditions to conveyor belt operations presents opportunities for technological advancement.
  • Multi-sensor Modalities: Integrating multiple sensor modalities and optimizing data acquisition protocols could reduce the amount of laboratory experiments required, though laboratory validation will remain essential.
  • Comprehensive Mineral Assemblage Datasets: A critical need is the development of comprehensive mineral assemblage training datasets that capture spectral interactions within whole rock samples. Such datasets would enable more robust validation and advance the understanding of automated classification systems.
  • Relative Mineral Amount Measurement: Measuring the relative amounts of different minerals more precisely could enhance the analysis and improve classification accuracy.

7.3. Personal Insights & Critique

This paper presents a very sensible and practical approach to a challenging problem. The integration of domain expertise with deep learning is a strength, particularly in fields like geology where empirical rules and expert knowledge are well-established but often difficult to codify for automated systems. The choice of 1D-CNN for mineral identification is appropriate given its proven effectiveness with spectral data.

Strengths:

• Hybrid Approach: The most significant strength is the intelligent hybrid architecture. It leverages the strengths of deep learning for low-level feature extraction (mineral identification) and expert systems for high-level reasoning and contextual understanding (rock classification based on assemblages and proportions). This is a strong model for knowledge-infused AI.
• Addressing a Key Gap: Directly tackling the mineral-to-rock classification gap is highly valuable, as it unlocks potential for automated material characterization beyond simple mineral detection.
• Uncertainty Handling: The inclusion of confidence and dominance thresholds for rock classification adds practical robustness, allowing the system to flag ambiguous cases rather than making potentially erroneous confident predictions.
• Geologically-Informed Data Strategy: The emphasis on geologically-informed sampling rather than generic data augmentation is crucial for developing models that are relevant and reliable in real-world geological contexts.

Potential Issues/Areas for Improvement (Critique):

• Limited Rock Sample Size: The most critical limitation is the extremely small dataset of 30 rock samples for evaluation. While the paper acknowledges this, it means the rock classification results, particularly the precision and recall figures, might not be representative or generalizable. The variable performance and misclassification patterns observed could be more systematically studied with a larger, more diverse rock dataset.
• Transparency of Expert Rules: While the knowledge base formulation is provided, the specific, detailed rules (beyond Table 1) for each rock type's mineral assemblage and proportions are not fully elaborated in the main text. This makes it challenging to fully understand the expert system's decision-making logic without accessing the code or further documentation.
• Handling of "Other" Category: The uncertainty-aware model's tendency to classify more samples as "Other" or misclassify them into other categories (e.g., increased granite misclassifications for sandstone samples) indicates a trade-off. While intended for robustness, it highlights the challenge of defining and learning an "unknown" class effectively in complex geological systems.
• Spectral Interactions in Whole Rocks: The paper mentions the need for datasets that capture spectral interactions within whole rock samples. This is a significant challenge, as Raman spectra from single minerals might differ when those minerals are part of a complex polymineralic rock due to matrix effects, mineral interfaces, or varying crystallographic orientations. The current approach assumes that single mineral spectra can be adequately combined to represent a rock, which might be an unverified assumption in all cases.

Transferability and Future Value: The methodological framework of knowledge-enhanced deep learning is highly transferable. This approach could be applied to other domains where:

1. High-accuracy component identification is achievable (e.g., using spectroscopy, microscopy, sensor data).

2. Complex assemblages or mixtures of these components define a higher-level classification.

3. Significant domain expertise exists in the form of rules, heuristics, or compositional guidelines.

   Examples include:

• Material Science: Classification of alloys, ceramics, or composite materials based on their elemental composition or phase mixtures.

• Environmental Monitoring: Identifying pollutants or biological agents in water/air samples based on spectroscopic signatures and known chemical combinations.

• Food Science: Classifying food products based on their chemical components and established recipes or quality standards.

  Overall, despite the limitations regarding dataset size, this paper provides a valuable and robust methodological blueprint for advancing automated material characterization by effectively bridging the gap between low-level data-driven spectral analysis and high-level domain-expert knowledge. The insights into uncertainty handling and the challenges of compositionally similar materials are particularly pertinent for future research in AI for scientific discovery and industrial automation.