Artificial intelligence and machine learning

Machine learning-based mineralogical analysis using planimetric method

M. G. Shishaev, V. V. Dikovitsky

Putilov Institute for Informatics and Mathematical Modeling of the Kola Science Centre of the Russian Academy of Sciences, Apatity, Russia

Abstract: The problem of automation of mineralogical analysis by the planimetric method is considered using the analysis of various types of apatite ores typical of the Khibiny deposit as an example.
Purpose: To study the efficiency of machine learning as a means for forming a features vector and solving the classification problem in various formulations during planimetric analysis of minerals.
Results: The study revealed the features of planimetric analysis as a classification problem, the properties of adjacency and homogeneity of classes in the space of identifying features are determined. Potential systematic errors in determining the content of the valuable component by the planimetric method are analyzed for various formulations of the classification problems. Experiments confirmed the possibility of directly using the pre-trained convolutional network ResNet-18 for forming a feature vector of classification objects, ensuring good separability of classes. Using the example of the ores under consideration, the high efficiency (over $98\%$ precision) of the neural network classifier and vectorizer ResNet-18 for identifying image elements related to the pure classes "apatite" / "non-apatite" is experimentally confirmed. High classification precision is maintained with planimetric grid cell sizes down to $2\times2$ pixels ($78\%$), and approaches $100\%$ at $20\times20$ pixels.
The effectiveness of a neural network approach to determining the specific grade of a valuable component in ore was studied. Experiments did not confirm the effectiveness of implementing planimetric analysis as a soft classification problem without significant modifications to the classifier’s architecture. However, they demonstrated the high effectiveness of the approach in a multi-class setting of the problem. The absolute error in determining the grade of a valuable component in the latter case depends on the number of classes and the ore type and, in the worst case, does not exceed $6\%$, which is higher than the accuracy of expert assessments by experienced mine geologists.
Practical relevance: the approach is applicable to the development of inexpensive, fast, and efficient express ore analyzers that do not require specialized equipment.

Key words and phrases: planimetric mineralogical analysis, machine learning, ResNet18, classification.

UDC: 004.932.2+004.89: 549.08
BBK: 32.973.26-018.2: 26.31

MSC: Primary 68T45; Secondary 00A99

Received: 12.10.2025
Accepted: 29.10.2025

DOI: 10.25209/2079-3316-2025-16-4-241-266