Artificial Neural Networks for Mineral Production Forecasting in the In Situ Leaching Process: Uranium Case Study
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper presents an approach to modeling uranium extraction using ANN. The authors are suggested to use clearer headings and subheadings to guide the reader through the methodology, results, and discussions.
The authors are advised to improve the literature review. It is better to discuss existing machine learning applications in ISL and how this approach differs or improves upon these methods.
It would be beneficial to provide the specifics of the ANN architecture used, including the number of layers, nodes, activation functions, and training algorithms. The selection of the input parameters should also be provided. Please consult “Machine learning enabled assessment of the vacuum freeze-drying of the kiwifruit” for input parameters selection, hyper-parameters tuning, correlational analysis, and other details regarding the ANN development.
It is recommended to provide more detail on how this data was generated, including any assumptions made and the range of parameters used.
The authors should elaborate the feature selection and any preprocessing steps taken to prepare the data.
Clarify the training and validation processes. Include information about the dataset split, cross-validation methods employed, and metrics used to assess model performance (e.g., RMSE, R²).
While the results indicate that the ANN achieves a high degree of accuracy, it would be useful to include a more detailed comparative analysis. Present quantitative results such as accuracy metrics alongside visualizations (e.g., plots of predicted vs. actual values).
Discuss the robustness of the model and its ability to generalize to different scenarios. Consider testing the ANN on datasets that were not part of the training set to validate its applicability.
Include a sensitivity analysis to assess how changes in input features affect the output of the ANN. This would provide insights into which features are most influential in predicting extraction rates.
Please highlight any limitations in this study, such as the size and scope of the training dataset or potential biases in the model. Discuss how these limitations affect the applicability of the findings. Please suggest directions for future research, including exploring more complex neural network architectures, integrating additional features, or applying the model to other mineral extraction processes.
Please ensure that terms are used consistently throughout the paper. For example, clarify the difference between “extraction rate” and “extraction degree” if they are intended to convey different concepts.
Comments on the Quality of English Language
Minor editing of English language required.
Author Response
Thank you for your valuable remarks. As per each point of the review, here are the amendments that were made to the article:
It would be beneficial to provide the specifics of the ANN architecture used, including the number of layers, nodes, activation functions, and training algorithms. The selection of the input parameters should also be provided. Please consult “Machine learning enabled assessment of the vacuum freeze-drying of the kiwifruit” for input parameters selection, hyper-parameters tuning, correlational analysis, and other details regarding the ANN development.
Thank you for this feedback. An additional information in regards the neural network architecture has been added into the article before Figure 3. Input parameters were selected as per the [https://doi.org/10.3390/min12111340], where the influence of each parameter has been investigated. This work is, in a way, a continuation of mentioned study. Current article explores the applicability of ANN into forecasting uranium recovery. Future work will include an additional study on hyperparameter tuning, with more features, as well as on real deposit production data available in mining enterprises.
It is recommended to provide more detail on how this data was generated, including any assumptions made and the range of parameters used.
The information in regards on how the training data was generated is described in Section 2. A synthetic dataset has been created based on the reactive transport modeling of In-Situ Leaching. Previously, by the authors a software complex has been developed and integrated in the subsidiaries of JSC NAC Kazatomprom, the largest uranium mining company in the World by production. Authors have developed 4 modules: geostatistical module to interpolate well data (lithology, mineralogy, filtration properties, etc.), CFD module to simulate the hydrodynamics of the process, chemical module to model the kinetics of the leaching process, as well as the economical module to optimize the process. The main complaint, however, by the geotechnologists on site, was the computational time. For one technological block, simulation of the lifecycle of which can take up to a day. While this is not an issue with currently exploited blocks, employing large computational software complexes to “quick and dirty” forecasting of probable production scenarios at pre-development stage is time consuming and inconvenient. The proposition of the geotechnologists was to apply neural modeling techniques to achieve probable extraction curves based on preliminary data alone, bypassing the costly, from computational standpoint, steps. Therefore, the reactive transport simulation functionality developed by the authors and described in [9], [4] and [14, 15], has been used to simulate the ISL process, obtain extraction curves and train the ANN to get these curves. The ISL is simulated not on the deposit, however, but on the laboratory equipment setting. This setting was previously used by the authors to obtain the reaction rate constants that directly impact the leaching process. The reactive transport model was the validated on real world deposits, with acceptable results. For ANN training, however, so far, a number of assumptions were made: the is 1D, permeability is homogeneous throughout the domain, pore clogging is not taken into account, concentration of reagent in leaching solution is constant over time. All these assumptions correspond to the empirical configuration. That is, the medium is indeed crushed and the sulfuric acid concentration is controlled to be of the same value. Only few important parameters were used in current study: production time, reagent concentration in the leaching solution, crystallized uranium concentration, flow velocity of the leaching solution, and the expected concentration of extracted uranium. In reality, more parameters influence the ISL: including, well locations, well screen depth, application additional reagents to expedite the leaching (pyrolusite, peroxide), etc. The ranges of parameters are shown in Table 2.
The authors should elaborate the feature selection and any preprocessing steps taken to prepare the data.
Uranium mining with ISL technique involves injecting a leaching solution, usually containing sulfuric acid (around 20 grams per liter), into the subsoil through the set of boreholes. Eventual pregnant solution (with dissolved mineral) is extracted via production well. At the site, geotechnologists control flow rates and acidity of the leaching solution. More reagents can be involved in the production to intensify the process. Well locations and well screens (through which the solution is injected from well into the subterrain) are also controllable parameters. Eventual concentration of uranium in pregnant solution is measured in production well. In order to simulate the process of ISL using reactive transport modeling, the data is always preprocessed. Based on flow rates, flow velocity is determined. All chemical components’ concentration is converted into mol per liter unit. However, for this experiment particularly, the question is, if we know flow velocity, concentration of sulfuric acid and solid uranium concentration would it be possible to forecast the production of uranium with some degree of error as compared to conventional, computationally costly methods (geostatistics, CFD, kinetics simulation), as quickly as possible, involving artificial neural networks. Therefore, already preprocessed data for reactive transport simulation was also used to train the ANN.
Clarify the training and validation processes. Include information about the dataset split, cross-validation methods employed, and metrics used to assess model performance (e.g., RMSE, R²).
Thank you for this valuable remark. An additional 216 test cases, not used while training the model were generated and evaluated for cross validation purposes. Values are both inside and outside the training range for input parameters.
The model should capture “correct” pattern during the training process, to be able to provide acceptable results on a data it was not trained on. The dataset was therefore split into two subsets: for training purposes and for validation. The dataset the model was tested on was “unseen” by the model during the training process. That is, initially 122 cases for training and 3 cases for cross validation of the model. The metrics that have been employed to assess the model were: NRMSD between the predicted and simulated curves for each time step, as well as the direct comparison of integrals between the predicted and simulated curves (which is basically a sum of mineral that would be extracted as calculated by neural model and reactive transport model).
While the results indicate that the ANN achieves a high degree of accuracy, it would be useful to include a more detailed comparative analysis. Present quantitative results such as accuracy metrics alongside visualizations (e.g., plots of predicted vs. actual values).
An additional figure with plots of percentage difference of the sum of uranium to be extracted for additional 216 test cases has been added into the article.
Discuss the robustness of the model and its ability to generalize to different scenarios. Consider testing the ANN on datasets that were not part of the training set to validate its applicability.
To 3 tests cases that were not part of the training process, an additional set of 216 test cases has been added.
Include a sensitivity analysis to assess how changes in input features affect the output of the ANN. This would provide insights into which features are most influential in predicting extraction rates.
A chart with values of each input parameter vs prediction error (total sum of extracted useful component) has been added into the article.
Please highlight any limitations in this study, such as the size and scope of the training dataset or potential biases in the model. Discuss how these limitations affect the applicability of the findings. Please suggest directions for future research, including exploring more complex neural network architectures, integrating additional features, or applying the model to other mineral extraction processes.
Conclusion section has been extended to reflect the points mentioned in this remark.
Please ensure that terms are used consistently throughout the paper. For example, clarify the difference between “extraction rate” and “extraction degree” if they are intended to convey different concepts.
Consistency of terminology usage of these notions has been fixed.
Author Response File: Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsThe manuscript describes and discusses an interesting AI-based idea for forecasting the uranium extraction during in-situ leaching processes. The conceptualized idea is implemented on a case study from Kazakhstan, where the relevance of the proposed approach is well demonstrated.
The manuscript is well structured and well written with good English quality. The results are clearly discussed and practical insight is provided into mining engineering for forecasting the uranium extraction during in-situ leaching processes. However, a number of moderate flaws/ambiguities still need to be addressed before the manuscript can be accepted for publication.
1. There is no description of the methodology and theoretical background of ANN model. It is important that authors address this issue because not all readers are necessarily experts in machine learning and data mining.
2. The authors need to clearly and logically state the criteria for configuring the ANN architecture (i.e., adopted network depth and neuron distribution). This will help the readers to reproduce the effective results for different real-case experiments.
3. There is a critical question about the splitting the training and testing datasets. It is common in machine learning applications that about 25-30% of the learning data are adopted as testing dataset to ensure the generalizability of the trained model. However, in this study, only about 2% of the learning data (i.e., 3 case out of 125 available case) are selected as testing cases. This decision may critically undermine the validity/generalizability of the trained model to predict new cases. Therefore, it is important for the authors to provide appropriate evidence to justify this choice.
4. The authors state in the caption of Figure 5 that the number of epochs is equal to 100. While the model appears to be trained over 780 epochs!
5. I recommend that the authors provide information about the code/software program used to implement the experiments and execute the ANN model.
Author Response
Thank you for your valuable remarks. As per each point of the review, here are the amendments that were made to the article:
- There is no description of the methodology and theoretical background of ANN model. It is important that authors address this issue because not all readers are necessarily experts in machine learning and data mining.
An additional context for the usage of ANN models was added into the introduction.
- The authors need to clearly and logically state the criteria for configuring the ANN architecture (i.e., adopted network depth and neuron distribution). This will help the readers to reproduce the effective results for different real-case experiments.
Neural architecture is described before Figure 3. The problem under consideration is inherently non-linear, necessitating the use of two hidden layers in the neural network to effectively capture the non-linearity. The selection of the number of neurons in these layers was guided by the relatively low number of features and the size of the available dataset. Expanding the dataset and incorporating additional parameters could require increasing the number of neurons to accurately capture complex patterns. However, this adjustment would likely lead to a trade-off, as a larger neural network would increase computational time and resource demands. Description of the neural architecture was extended.
- There is a critical question about the splitting the training and testing datasets. It is common in machine learning applications that about 25-30% of the learning data are adopted as testing dataset to ensure the generalizability of the trained model. However, in this study, only about 2% of the learning data (i.e., 3 case out of 125 available case) are selected as testing cases. This decision may critically undermine the validity/generalizability of the trained model to predict new cases. Therefore, it is important for the authors to provide appropriate evidence to justify this choice.
Thank you for this valuable remark. An additional test set, not utilized during the training phase of the model, was generated and evaluated for cross-validation purposes. The input parameter range was also extended beyond the scope of the training dataset, covering flow velocities from 0.45 to 0.95 meters per day, reagent concentrations from 5 to 35 grams per liter, and solid uranium mass concentrations ranging from 0.03% to 0.06%. In total, 216 additional test cases were evaluated. The model demonstrates robustness with low deviations when the input parameters fall within the training dataset's range. However, significant deviations are observed for reagent concentrations at 5 grams per liter, which is below the training minimum of 10 grams per liter. Additionally, errors are noted for solid uranium content below 0.00035 kg.kg-1, which was the minimum used in the training set. No consistent error patterns were identified for flow velocity values, likely due to the gradual variation of these values within the extended range. These limitations were also added into conclusion section.
- The authors state in the caption of Figure 5 that the number of epochs is equal to 100. While the model appears to be trained over 780 epochs!
The caption has been fixed. Thank you for noticing this error.
- I recommend that the authors provide information about the code/software program used to implement the experiments and execute the ANN model.
An additional information on the technology stack and versions used to develop this program has been added into the introduction.
Author Response File: Author Response.docx
Reviewer 3 Report
Comments and Suggestions for AuthorsThe subject of the submitted manuscript concerns significant issues of the use of artificial neural networks (ANN) in modeling the uranium In-Situ Leaching (ISL) process. The results presented in the manuscript prove that the regression type neural networks can be applied to model the recovery rate of In-Situ Leaching process of mineral production. The important thing for ISL simulation is that all intermediate, computationally expensive steps have been omitted while ANN has been trained to provide extraction curves. This approach to the topic does not appear in other published materials. The title and summary correspond to the content of the manuscript. The literature review is current (more than half of the cited items are from the last 5 years), although not very extensive. The references are appropriate. The authors may consider supplementing the list of literature with the following items:
1. Ion migration in in-situ leaching (ISL) of uranium: Field trial and reactive transport modelling. Journal of Hydrology 2022, https://doi.org/10.1016/j.jhydrol.2022.128634
2. Synthetic data generation for ANN modeling of the hydrodynamic processes of in-situ leaching. Scientific Journal of Astana IT University 2024, https://doi.org/10.37943/17STXF5228
The use of artificial neural networks in modeling has been described clearly and sufficiently. Regarding the methodology, authors should take into account the comments presented below. Analysis of calculation results is correct. The conclusions are consistent with the presented calculation results and refer to the main question posed because omitting all intermediate steps did not worsen the prediction results. The conclusions take into account future research.
I have only two comments on the manuscript:
- P.7, Table 4 - no explanation why the parameter values ​​given in the table were adopted,
- P.10, Fig.10 - poor readability of the drawing.
Author Response
Thank you for your valuable remarks. As per each point of the review, here are the amendments that were made to the article:
- Ion migration in in-situ leaching (ISL) of uranium: Field trial and reactive transport modelling. Journal of Hydrology 2022, https://doi.org/10.1016/j.jhydrol.2022.128634
- Synthetic data generation for ANN modeling of the hydrodynamic processes of in-situ leaching. Scientific Journal of Astana IT University 2024, https://doi.org/10.37943/17STXF5228
The use of artificial neural networks in modeling has been described clearly and sufficiently. Regarding the methodology, authors should take into account the comments presented below. Analysis of calculation results is correct. The conclusions are consistent with the presented calculation results and refer to the main question posed because omitting all intermediate steps did not worsen the prediction results. The conclusions take into account future research.
I have only two comments on the manuscript:
- P.7, Table 4 - no explanation why the parameter values ​​given in the table were adopted,
- P.10, Fig.10 - poor readability of the drawing.
An additional information has been added in regards of specific parameters being chosen for this particular study has been added into the article.
An introduction has been extended to reflect valuable information from the first article suggested. The second suggested article, however, is directed at grid data preparation for Physics-Informed Neural Networks. The main idea of current study, is to explore the applicability to quickly forecast probable uranium production by bypassing intermediate steps such as geostatistical interpolation, CFD simulation, as well as modeling of the kinetics of chemical processes of ISL.
Figure 10 has been resized to increase readability.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe manuscript has been improved significantly and can be considered for publication.