As part of the BRidge Alfa R&D project, components of the modeling and forecasting system for the distribution of copper content in 3D space are being developed. The project involves analyzing the necessary data required for industrial research in the project. Ontological and semantic requirements have been prepared, and through cooperation with KGHM Polska Miedź S.A., research was conducted on drilling holes with sampled cores, samples from mining workings, and data for determining the elevation coordinate of the sandstone roof (h), based on leveling points, the 3D isocline map of the sandstone roof, and the model in the form of a mesh grid created based on the isoclines of the sandstone roof. Blender software was used, and its capabilities, along with optional extensions, proved suitable for presenting irregular three-dimensional maps. PostGIS and QGIS were utilized for spatial analysis tools. In the preliminary research, spatial data visualization and analysis were carried out. Once the data was prepared, pattern extraction and grouping identification began. The purpose of data clustering is to identify subareas with similar geological conditions while also highlighting the differences between subareas. It is important to uncover correlations or co-occurrence of specific features in each group. As part of industrial research, spatial data clustering was performed, where numerical data was standardized, and categorical data was separated into multiple dichotomous variables. Custom software was developed in Python, implementing various machine learning models to replicate dependencies in the analyzed data. The results of different versions of the models were compared, and data visualization methods were used in the analysis. A key task at this stage was developing quality measures for models in regression, clustering, and classification problems. Software tools were also created to generate maps displaying copper distribution using various interpolation methods, such as linear, quadratic, spline, and simple kriging. Modified kriging models were based on appropriately selected test data, testing kriging for different scenarios simulating mining progress and various numerical methods. Kriging was also tested in combination with regression models and machine learning models. Modified statistical methods and a hybrid approach, combining different models and methods, allowed for the planning and execution of experiments evaluating the newly developed technology. The aim of these experiments was to find the best methods for predicting copper content in mining areas represented by a set of three-dimensional points containing information on location, rock type, copper content, and the time of sample collection. Several benchmark models were used, including linear interpolation, kriging, linear regression, support vector regression (SVR), random forests, artificial neural networks (ANN), kriging + linear regression, kriging + SVR, kriging + random forest, and kriging + ANN. All models were implemented in a way that allowed easy modification of hyperparameters and repeated experiments. The quality of the results was initially assessed using the R2 coefficient, and later D0, D1, and D2 accuracy indicators were calculated for predicting copper content within allowable error ranges of zero, one, or two classes of copper content. The results varied depending on the area and experiment. Using the height above the sandstone level significantly improved the results compared to earlier trials with different heights. Kriging with regression in the form of rectangular prisms yielded promising results, with a restriction on overlapping areas in the test and validation data. For rectangular prism data, standalone regression models also proved effective. Adding a minimum sample number restriction to create a block improved the predictions of most models. The prediction of original samples (before aggregation) gave comparable results across different forms of test data used. Three series of experimental research were conducted as part of the industrial research. Models prepared in each series returned predictions as real numbers, which were then discretized into one of five copper resource classes. The DN measure represents classification accuracy with an allowable error of N classes. Test results showed that the best fit was achieved with a random forest model combined with kriging. After generating the test set and training models based on the combined training and validation sets, the final fitting of the best model was: D1 = 0.87, D2 = 0.98. As part of the development work, technology components integrating the developed models with functional and non-functional components were created, enabling the use of the solution in the planning process of resource extraction. The components developed in industrial research for data analysis and visualization were integrated with new components including an API responsible for batch and incremental acquisition of data from mining databases or planning and production applications. To demonstrate the results obtained by different models, a prototype application was developed, incorporating machine learning components. Functional components were designed and implemented for visualizing actual and predicted samples generated by machine learning models. The application allows the user to analyze detailed information on the chemical elements in the samples, in this case, copper content. The application was implemented as a web-based one-page application, incorporating the entire functionality available to the user. Python Dash, machine learning scripts, and a PostgreSQL database server (with PostGIS extension) were used for the prototype. The application was developed in Flask using Python, with peripheral technologies such as Dash and Plotly, and OpenStreetMap maps were also used. The application managed data using the Pandas and Numpy libraries, and Scikit-Learn and Pykrige were used for machine learning processes. Visualization data was obtained through machine learning scripts communicating with the PostgreSQL database. At the business logic level, the application loaded these data using Pandas and created queries for communication with the presentation layer. The interface is supported by the Dash library, integrated with Flask and Plotly. An interface for sharing forecast data, considering model aspects and prediction times, was also developed to allow analysis and visualization of these data in peripheral software used in mining companies. These integration components were implemented using Flask RestAPI, completing the technology stack. The main objective of the extraction simulation is to verify the assumptions of the extraction plan. This verification was conducted on new data, divided into subsets of all samples grouped chronologically to fully reflect the mining process. The simulation utilized all available information, particularly data from borehole samples that helped group data. These data were used to train all four models, including the artificial neural network model. Based on the above, a simulation of the resource extraction process was planned and executed, using updated machine learning models and incremental data sets for iterative experiments to fully reflect the forecasting process.

The provision of the complete application prototype served as a demonstration of the final form of the technology and the achievement of Technology Readiness Level (TRL) 8. By verifying the mining plan through the simulation of the exploitation process, it was confirmed that the technology can be applied under the expected conditions.