Advancements in machine learning applications for mineral prospecting and geophysical inversion: A review

2.2.1.1 Data enhancement strategies

Data augmentation is a widely used technique in DL that enhances the diversity of data through methods such as image rotation, scaling, and cropping. Combining different DL networks with data augmentation techniques, especially in the case of data scarcity, has yielded good research results [16]. However, geological exploration data possess directional and spatial relationships. For example, rotation operations may change the direction of a fault and lead to absurd geological phenomena and prediction results. This outcome does not comply with the core principle of data enhancement – does not change the semantics of the labels. Some researchers have generated sufficient samples by randomly discarding data enhancement methods on the basis of CNNs, and some have increased the amount of available data by conducting clipping to ensure that geological spatial information is not destroyed. The research results show that these methods have certain effects. Similar data enhancement methods include sliding windows and injected noise, which make small spatial adjustments to the original image data, but an increase in the number of samples may lead to data redundancy and thus overfitting.

Conditional GANs (CGANs) are powerful generative models that can create additional training samples when data are limited, thereby enhancing the generalizability of the constructed model. In geological exploration scenarios, CGANs can learn from various datasets and generate feature samples without distorting geological characteristics or altering their spatial significance. Wu et al. [17] used the idea of a generative model and applied a CGAN for data enhancement to solve the small sample size problem, and they compared the sliding window technique with their new approach; the results showed that the data enhancement strategy based on the CGAN had stronger anti-overfitting capabilities. AE technology has been widely used in the field of geochemistry. The basic idea of an AE is to encode and reconstruct the sample data during the training process and use the induced reconstruction error to optimize the model parameters in reverse. However, owing to the low contribution rates of samples with low probabilities in AE networks, these samples are easily recognized as abnormal samples by trained models. The features learned by an AE may only be the identity representation of the original input, and the essential features of the sample cannot be extracted. Compared with AEs, GANs have a greater ability to generate more realistic and high-quality data. Xie et al. [18] proposed an AEGAN model, which embeds an AE network into an adversarial network structure, and built a stable and high-quality generation model by combining the characteristics of the AE and GAN. This approach compensates for the shortcomings of AEs and GANs.

2.2.1.2 Transfer learning strategies

In addition to data augmentation, transfer learning is another approach for solving problems with few samples. The basic idea of transfer learning is to utilize an existing model (usually one that has been trained on a large-scale dataset) to solve a new problem. This approach accelerates the learning process and improves the performance of the constructed model on the new task, especially if the amount of data available for the new task is small, and it can improve the performance of the model in scenarios with limited resources. Huang et al. [19] proposed a DL framework for selective knowledge transfer that captures features at different scales via inflated convolution and selectively transfers the weight parameters of the source network, which helps improve the performance of the developed model without adding additional computational costs. A masking technique was also introduced to maintain the weights between the associated mineral elements and the main elements. And Li et al. [20] proposed a 3D mineral prediction method based on transfer learning. Compared with traditional methods that can only predict the study area when information is abundant, transfer learning can still predict deep mineral prospects in cases with incomplete data. However, transfer learning usually requires experiments to select relevant and differentiated source domains and specific algorithms that contribute to the training process.

2.2.1.3 Self-supervised and semi-supervised learning strategies

Currently, MPM methods primarily employ supervised learning to train networks. However, a contradiction is present between the scarcity of mineral exploration data and the requirement of a substantial number of high-quality samples for supervised learning. Additionally, the collection and labeling of these samples demand significant human resources. Miao et al. [21] utilized self-supervised comparative learning to complete prediction tasks. Self-supervised learning is a ML approach that does not rely on externally labeled data but instead generates training signals from the structure and properties of the input data themselves. This mineral prediction method initially leverages many unlabeled exploration samples for comparative learning, enabling the extraction and understanding of data features. The learned features are then transferred to a downstream supervised learning network. Since the relationships and essential features have already been identified, the downstream network can make minor adjustments with a limited number of samples. By integrating self-supervised comparative learning, the sample size requirements of the network can be significantly reduced.

In addition, the Siamese network is a commonly used DL framework for similarity metrics and comparison tasks and is often employed for image recognition and text similarity analyses. The network structure is characterized by two identical subnetworks that share a common set of parameters and are able to simultaneously process two input samples and evaluate the similarity between the samples by comparing their feature representations. In this way, the similarities and differences between the samples can be learned to distinguish the similarities and differences between unseen objects. Ding et al. [22] used the Siamese network architecture for mine prediction; because this network can distinguish between different data, it can alleviate the insufficient sample problem encountered in mineral exploration scenarios to a certain extent. The experimental results showed that the developed model achieved better performance than a CNN model alone in finding mines.

2.2.2.1 Attention mechanism-based strategies

The use of attention mechanisms in DL allows models to concentrate on the salient aspects of the input data, thereby enhancing the ability to extract pivotal feature channels from intricate datasets. Gao et al. [23] proposed the multiscale feature attention framework DL model, which uses convolution kernels with different sizes to generate more labeled sample data and uses channel and spatial attention mechanisms to assign different weights to geological image feature data. Xu et al. [24] proposed the Geo-DCNN, which is based on a deep CNN, and added a pyramid feature extraction structure and a multiscale classification structure to this model, as well as an attention mechanism for identifying the spatial distribution and channel weights; these adjustments improved the feature extraction ability and recognition efficiency of the model. Liu et al. [25] fused a 3D CNN with a CBAM to generate a 3D CNN-based MPM model with an attention mechanism. This model integrates the integrated attention mechanism of the CBAM into CNNs to enhance the overall performance of the network. The verification results showed that the CNN + CBAM model could focus on the target object and capture high-dimensional ore control characteristics. The model was more accurate in terms of predicting deep drilling data, and the exploration results were highly consistent with the real values or spatial distribution characteristics of known voxels.

The attention mechanism, as a specially designed network structure that mimics human visual perception, enables the model to automatically focus on key features in the input while ignoring irrelevant details. By integrating the spatial attention module and the channel attention module, the mechanism significantly improves the model's ability to identify and emphasize key geological features in mineralization prediction. In the application of multiphysics modeling, the attention model realizes the efficient integration of multi-source data by dynamically capturing the dependencies and spatial heterogeneity of input data. This approach effectively compensates for the shortcomings of the traditional MPM approach, which tends to ignore the correlation between evidence layers when fusing different evidence layers to generate mineral potential maps, resulting in the loss of useful information. The attention mechanism not only enhances the accuracy of prediction, but also provides a more robust and efficient solution for mineralization prediction [26].

2.2.2.2 Multinetwork combinations and multiscale strategies

In the field of DL, CNNs are popular because of their excellent performance in image recognition and processing tasks. However, the traditional CNN architectures have some limitations when addressing certain problems. Specifically, they typically extract features at a single scale, which restricts the capacity of the constructed model to capture multiscale information. This is a critical issue in mineralization studies, where mechanisms must be understood in the context of broader geological patterns. Yang et al. [27] improved upon GoogLeNet by introducing four sets of convolutional kernels to simultaneously extract and integrate features at different scales, thereby significantly improving the accuracy of their model. This approach allows the network to capture information from multiple scales in a single iteration, effectively enriching its feature representations.

The integration of neural networks is also a strategy for enhancing the generalizability of models [28]. Li et al. [29] proposed a novel network architecture that combines a transformer and a CNN: a CNN-transformer model. The transformer architecture is well-known for its ability to address long-range dependency problems, and it especially exhibits significant advantages in global feature extraction cases. However, it is relatively weak at capturing local details. In contrast, CNNs are strong local extraction methods but possess weak global capture capabilities. By combining the strengths of CNNs and transformers to form a CNN-transformer network, the resulting network is able to capture both local and global key information, thus achieving a more comprehensive feature understanding in image processing and recognition tasks. This innovative network design not only compensates for the shortcomings of single models but also opens new possibilities for the application of DL in complex data processing tasks involving geological and mineral exploration.

3.2.1.1 Joint inversion strategies

An underground model has many physical properties, and different physical inversion methods can infer the target underground structure according to these properties. For example, electromagnetic data can be used to infer subsurface resistivity, and seismic and gravity data can be employed to infer subsurface velocity and density conditions [57,58]. Therefore, more appropriate methods should be selected for different exploration environments [59]. In recent years, with the continuous accumulation of geophysical data, underground information has become increasingly diverse. The inversion of a single physical attribute does not consider the correlations and constraints among multiple underground attributes. To effectively use these physical observation data and give full play to the feature extraction capabilities of DL networks, joint inversion can be adopted [60–63].

As shown in Figure 3, the joint inversion schematic illustrates the process of subsurface model reconstruction through the combined training of signals with different geophysical attributes. As the joint inversion combines the characteristics of multiple observation data, it is able to realize mutual constraints and complementary information between the data in the training process. When the error of the training results is controlled within a reasonable range, the final prediction results are output; otherwise, a new round of iterative training will be entered to further optimize the model performance by adjusting the neural network parameters.

Figure 3

Joint inversion process.

For example, Wang et al. [64] combined seismic data and electromagnetic data to conduct joint inversion. Seismic data can provide a high-resolution underground velocity model, whereas electromagnetic data, although they provide resistivity data with low resolutions, are sensitive to underground anomalies. Therefore, the combination of these two types of data enables them to effectively complement and constrain each other. Bai et al. [65] designed a neural network based on a CNN for the inversion of gravity data and magnetic data, which are two physical properties that are often mutually verified and constrained in traditional methods. Hu et al. [62] proposed a DL-enhanced framework for joint multiphysics field inversion that uses DL networks as a bridge connecting individual inversions to achieve information extraction and fusion. Guo et al. [59] proposed jointly inverting MT data and seismic data in combination with DL constraints. The key to this method is that it attempts to understand the interrelationships between the resistivity and velocity of the Earth's interior and their structural characteristics through DL models. Through this method, researchers are able to more accurately reconstruct the structure of the Earth’s interior. A better understanding of the relationship between resistivity and velocity is thus attained.

Data with different attributes can be combined, and multicomponent electromagnetic data with different sizes can also be used for joint inversion. Magnetic fields have multiscale characteristics; in other words, the magnetic fields observed at different altitudes can represent different scales. Shi et al. [66] combined multitype and multiscale magnetic tensor data and magnetic three-component data for inversion purposes, which helped the researchers conduct debugging between different scales and different types of magnetic data and reduced the number of solutions yielded by the inversion results. Some of the above joint inversion methods were carried out through pure DL, and some were assisted by DL to associate various independent physical observation data. Compared with independent physical attribute inversion, joint inversion has a higher data utilization rate and produces better inversion results. Joint inversion is an inevitable research direction for the future.

3.2.1.2 Self-supervised and semi-supervised learning strategies

As a result, mineral exploration analysts have begun to look at another DL strategy, self-supervised learning, as a way to solve the sample scarcity problem. Self-supervised learning can use unlabeled data to learn, extract the operation rules and main features of tasks, and then fine-tune the constructed model on a small amount of labeled data [67]. In this way, inversion can be carried out in mineral exploration scenarios with samples that lack labels and small amounts of subsurface real data. For example, Li et al. [68] proposed a self-supervised 3D gravity inversion method, which directly learns the data of the target exploration site through a closed loop consisting of an inversion model and a forward model so that the law of the mineral exploration data themselves can be utilized instead of relying on manually labeled training data. This method is suitable not only for 3D gravity inversion but also for other physical inversions. For example, Li et al. [69] subsequently applied this method to 3D magnetic field inversion with slight modifications. In the absence of real data, some researchers have used GANs to reduce the reliance of the existing methods on labeled samples. For example, Qiao et al. [70] designed a CycleGAN based on the inversion of gravity anomalies; this approach can generate data that are close to those of the initial sample with a small amount of prior information so that a small number of training samples can be used to obtain good prediction results. These experiments and studies have proven the feasibility of unsupervised learning and semisupervised learning (GANs are semisupervised learning techniques) for performing inversions when data are lacking [71,72]; this is also a direction for future research.

3.2.2.1 Strategies that add constraints and prior information

A nonpure data-driven inversion method integrates physical constraints with prior information and expert knowledge into a neural network. Such constraints can enable the network to learn the inherent physical laws exhibited by the inversion process during training to more reasonably guide the network parameter updates [42]. For example, if the observed resistivity data have vertical variation characteristics, when the same anomaly is located at different vertical locations, the apparent resistivity data will present large differences, which will lead to ambiguity during the DL inversion process [73–76]. Liu et al. [77] incorporated layered information into their network to address this phenomenon, and close correlations were observed between the data of different strata and the geological anomalies present at the corresponding depths. Therefore, when analyzing apparent resistivity data, formation information can be regarded as an effective supplement to the depth information processed by a neural network. This approach enhances the ability of the utilized model to identify the details of underground structures, thereby improving the accuracy of geological anomaly detection. Ling et al. [78] introduced a combined data-driven and physically driven weighted loss function to make their network consider the depth factor, which alleviated the dependence of the network on the training set range to a certain extent. In addition, since the anomalous area of interest occupies only a small part of the total underground space, a model may be biased to fit more background data (because the amount of background data is larger), resulting in inaccurate anomalous body predictions. In this context, Jiao et al. [42] designed a multiconstrained UNet++ network for magnetic data inversion, in which the Dice coefficient and the previous fit constraint terms were integrated into the loss function. The core idea of the Dice coefficient is to guide the model to focus on the accurate location of the anomalous body source while avoiding the limitation of the local minimum value during the optimization process of the loss function. The constraint item of the previous item is used to calculate the difference between the magnetic field anomaly data obtained after simulating the 3D model predicted by the network in the forward direction and the actual observed data to ensure the consistency of the predicted results with the actual geological conditions. In this way, the model is able to more effectively capture the key features contained in the data, which improves its overall predictive performance. Effectively integrating physical constraints and expert knowledge into DL models and building hybrid architectures based on them by combining multiple networks not only helps to improve the generalization ability of the models, but also significantly improves the inversion accuracy. However, the current difficulty lies in how to effectively integrate this geologic knowledge and reasonably fuse it with DL methods. This area still needs further in-depth research and exploration [79,80].

3.2.2.2 Noise removal strategies

The noise faced by mineral exploration methods can be widely divided into internal noise and external noise. The so-called internal noise is inherent in system hardware and equipment, whereas external noise includes man-made noise, natural noise, and the noise induced by movements [81–83]. The advantages of introducing physical knowledge and constraints, which can improve the degree of generalization of a network to a certain extent and provided it with more antinoise properties than those possessed by a pure data-driven approach, are discussed above. However, it is often difficult to obtain satisfactory results when such complex noise interferes with geophysical measurements [44]. Therefore, denoising inversion data is an important prerequisite for accurately performing inversion. For noise processing, the traditional method involves addressing all types of noise in a step-by-step manner, but these methods require many resources, and some parameters need to be set by relevant personnel. Although all types of noise can be effectively processed, the experience and knowledge reserves of relevant staff have high requirements, and large amounts of computing resources are needed.

In the inversion process of geophysical data, DL models have demonstrated their unique advantages in noise removal. By utilizing a variety of DL architectures, including artificial neural networks, CNNs, and self-coders, we are able to effectively remove noise from exploration data. Compared with traditional methods, these techniques can reduce the influence of manual subjective judgment on the inversion results, thus improving the objectivity and accuracy of data processing [83,84]. Utilizing DL to denoise noisy data before conducting inversion is the most direct method. For example, Zhang et al. [85] designed a deep residual network by first using a small time window to scan an MT time series and then by using Gramian angular field technology to assist the noise detection process so that the input time series could be effectively divided into noise-free and noise-containing data segments to form a training set; this training set was then input into the network for training. Han et al. [82] built two RNNs for denoising: one network for screening out noisy data and another network for denoising. The results of denoising greatly improve the subsequent inversion process. DL-based denoising can also be combined with traditional methods. For example, Wang et al. [86] combined a deep residual denoising CNN with shift-invariant sparse coding (SISC, which is a sparse representation technology that is used to further process and optimize denoised data) for denoising. Specifically, the DL part is responsible for automatically learning noise features from the input data for preliminary denoising. The denoising results are then transmitted to the SISC module for further fine processing to improve the quality of the denoised data.

All of the above denoising processes use DL to denoise geophysical detection data before performing inversion. Another strategy is to inject noise into the given measurement data and use an inversion neural network to automatically denoise and obtain the inversion results. Compared with conducting processing before inversion, this is an end-to-end method with continuity and unity. For example, Liu et al. [44] proposed three different noise injection strategies, including the addition of synthetic relative noise following a Gaussian distribution, a multiwindow filtering scheme, and a data enhancement method based on two strategies to simulate the noise situation in real measurement data. Notably, both conducting noise processing before inversion and the noise injection task implemented during inversion are single-task learning networks in DL. Among the available noise removal methods, a type of network called the multitask learning network can carry out noise removal and other tasks simultaneously [87]. For example, Liu et al. [88] proposed a multitask learning neural network structure based on a shared-layer transformer, which can simultaneously denoise and invert aeromagnetic data and uses shared features to mutually constrain the denoising and inversion tasks. Zhang et al. [89] designed a multitask learning framework based on an improved UNet structure for gravity data inversion and denoising. Through its multitask learning framework, this network can perform data inversion while conducting denoising, effectively improving the accuracy of the inversion results obtained for noisy data. The multitask learning strategy can indeed improve the accuracy of inversion, but it should be noted that the weights selected for different tasks affect the final result. Multiple tasks can bring more constraints and information but also introduce some biases; thus, highly correlated tasks tend to improve, whereas poorly correlated tasks tend to cause this approach to backfire.