Allan Erlikhman Medeiros Santos
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Allan Erlikhman Medeiros Santos has a total of 29 co-authors. Here are the top co-authors by number of shared publications:
- Denise de Fátima Santos da Silva
Universidade Federal de Minas Gerais
8 shared publications - Tatiana Barreto dos Santos
Universidade Federal de Ouro Preto, Brazil
4 shared publications - José Matheus Vieira Matos
Universidade Federal de Ouro Preto, Brazil
3 shared publications - Naim Khalil Ayache
Centro Federal de Educação Tecnológica de Minas Gerais
3 shared publications - Guilherme Alzamora Mendonça
Universidade Federal de Ouro Preto, Brazil
2 shared publications - Milene Sabino Lana
Federal University of Ouro Preto – UFOP, Brazil
2 shared publications
Latest Publications
- Open Access
Phase Diagram for Rock Mass Compactness Classification from Reverse-Circulation Drill Chip Granulometry
Rock mass compactness is critical for geotechnical characterization, but core logging is slow. This study develops an auditable framework to classify rock mass compactness (Compact, Semi-Compact and Friable) from the granulometry of reverse-circulation drill chips. Tyler particle-size distributions from intervals of an itabirite iron-ore deposit were projected by a supervised autoencoder onto a two-dimensional phase space in which multinomial logistic regression defines closed-form decision boundaries. Over 20 repetitions of leave-one-out cross-validation, the framework achieved accuracy, exceeding linear classifiers trained on the sieve fractions (–); a Random Forest reached but provides no planar representation. Withholding entire drill holes reduced accuracy to (baselines 18–); so, performance applies to the sampled holes. Y-randomization gave , the minimum attainable. In the archived reference model, encoder Jacobians showed that reflects the depletion of intermediate chips (#28–#14) and coarse-chip retention (#14 and #6), a reading specific to that model because the axes are not fully reproducible between training runs. The encoder reduces to spreadsheet-implementable matrix operations, and the site-calibrated framework requires validation before transfer to other deposits.Open AccessPhase Diagram for Rock Mass Compactness Classification from Reverse-Circulation Drill Chip Granulometry
Rock mass compactness is critical for geotechnical characterization, but core logging is slow. This study develops an auditable framework to classify rock mass compactness (Compact, Semi-Compact and Friable) from the granulometry of reverse-circulation drill chips. Tyler particle-size distributions from intervals of an itabirite iron-ore deposit were projected by a supervised autoencoder onto a two-dimensional phase space in which multinomial logistic regression defines closed-form decision boundaries. Over 20 repetitions of leave-one-out cross-validation, the framework achieved accuracy, exceeding linear classifiers trained on the sieve fractions (–); a Random Forest reached but provides no planar representation. Withholding entire drill holes reduced accuracy to (baselines 18–); so, performance applies to the sampled holes. Y-randomization gave , the minimum attainable. In the archived reference model, encoder Jacobians showed that reflects the depletion of intermediate chips (#28–#14) and coarse-chip retention (#14 and #6), a reading specific to that model because the axes are not fully reproducible between training runs. The encoder reduces to spreadsheet-implementable matrix operations, and the site-calibrated framework requires validation before transfer to other deposits. - Open Access
Rock Strength Classification in a Brazilian Iron Ore Mine Using Operational Drilling Variables and Machine Learning
Rock strength is a key parameter for mine planning and operational optimization, but conventional laboratory testing is costly and provides limited spatial coverage. This study develops a methodology for classifying operational rock-strength classes in a Brazilian iron ore mine using reverse circulation (RC) drilling data and machine learning. Approximately 700 m of drilling data from ten boreholes were analyzed using operational variables acquired by onboard sensors. Data consistency was assessed through twin-hole analysis using Principal Component Analysis (PCA) and LSTM-based autoencoders, while K-Nearest Neighbors (KNN) was used to assess the reproducibility of rule-based operational-state labels. Five supervised algorithms were subsequently evaluated for classifying four operational rock-strength classes assigned by depth correlation with adjacent diamond drillholes and supported by available UCS-based geotechnical records, rather than by direct UCS testing of each RC interval. Random Forest achieved the best internal performance, with a mean accuracy of 90.2% across 50 stratified random partitions, followed by SVM at 88.1%. External validation of Random Forest on an independent blind drillhole achieved 93.1% accuracy, correctly classifying 54 of 58 intervals. These results provide preliminary evidence of model transferability and highlight the potential of RC operational data for indirect rock-strength classification.
Open AccessRock Strength Classification in a Brazilian Iron Ore Mine Using Operational Drilling Variables and Machine Learning
Rock strength is a key parameter for mine planning and operational optimization, but conventional laboratory testing is costly and provides limited spatial coverage. This study develops a methodology for classifying operational rock-strength classes in a Brazilian iron ore mine using reverse circulation (RC) drilling data and machine learning. Approximately 700 m of drilling data from ten boreholes were analyzed using operational variables acquired by onboard sensors. Data consistency was assessed through twin-hole analysis using Principal Component Analysis (PCA) and LSTM-based autoencoders, while K-Nearest Neighbors (KNN) was used to assess the reproducibility of rule-based operational-state labels. Five supervised algorithms were subsequently evaluated for classifying four operational rock-strength classes assigned by depth correlation with adjacent diamond drillholes and supported by available UCS-based geotechnical records, rather than by direct UCS testing of each RC interval. Random Forest achieved the best internal performance, with a mean accuracy of 90.2% across 50 stratified random partitions, followed by SVM at 88.1%. External validation of Random Forest on an independent blind drillhole achieved 93.1% accuracy, correctly classifying 54 of 58 intervals. These results provide preliminary evidence of model transferability and highlight the potential of RC operational data for indirect rock-strength classification.
- Open Access
- 22 April 2026
- Published by Elsevier in SSRN Electronic Journal
Automated Reinforcement Learning with Bayesian Hyperparameter Optimization and Invalid Action Masking for the Sequential Ordering Problem
The Sequential Ordering Problem (SOP) presents a significant challenge in Combinatorial Optimization (CO) due to its strict precedence constraints, rendering exact methods computationally prohibitive for large-scale instances. While Reinforcement Learning (RL) offers a promising alternative, its effectiveness is often hindered by the difficulty of tuning interacting hyperparameters and the inefficiency of learning constraints through trial-and-error. This work proposes Masked AutoRL-SOP, a novel framework that synergizes the SARSA algorithm with a deterministic invalid action masking mechanism and Bayesian Hyperparameter Optimization (HPO) via the Tree-structured Parzen Estimator (TPE). The method removes invalid actions from the action space prior to action selection and jointly optimizes hyperparameters (learning rate, discount factor and exploration rate) to capture non-linear dependencies. Experimental results on 28 benchmark instances from TSPLIB repository demonstrate that the proposed framework matches or outperforms state-of-the-art Automated Reinforcement Learning (AutoRL) baselines in all instances. The method delivers substantial quantitative gains in efficiency, reducing runtime by an average of 81% across all instances and achieving speedups of up to 130 times on large-scale problems. The total experimentation time decreases from approximately one week to under eight hours, while the final model training phase consistently requires less than five seconds per instance. The study also provides empirical evidence that optimal hyperparameter configurations are topology-dependent and non-linear.
- 22 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAutomated Reinforcement Learning with Bayesian Hyperparameter Optimization and Invalid Action Masking for the Sequential Ordering Problem
The Sequential Ordering Problem (SOP) presents a significant challenge in Combinatorial Optimization (CO) due to its strict precedence constraints, rendering exact methods computationally prohibitive for large-scale instances. While Reinforcement Learning (RL) offers a promising alternative, its effectiveness is often hindered by the difficulty of tuning interacting hyperparameters and the inefficiency of learning constraints through trial-and-error. This work proposes Masked AutoRL-SOP, a novel framework that synergizes the SARSA algorithm with a deterministic invalid action masking mechanism and Bayesian Hyperparameter Optimization (HPO) via the Tree-structured Parzen Estimator (TPE). The method removes invalid actions from the action space prior to action selection and jointly optimizes hyperparameters (learning rate, discount factor and exploration rate) to capture non-linear dependencies. Experimental results on 28 benchmark instances from TSPLIB repository demonstrate that the proposed framework matches or outperforms state-of-the-art Automated Reinforcement Learning (AutoRL) baselines in all instances. The method delivers substantial quantitative gains in efficiency, reducing runtime by an average of 81% across all instances and achieving speedups of up to 130 times on large-scale problems. The total experimentation time decreases from approximately one week to under eight hours, while the final model training phase consistently requires less than five seconds per instance. The study also provides empirical evidence that optimal hyperparameter configurations are topology-dependent and non-linear.
- Open Access
- 5 January 2026
- Published by Universidade Federal do Rio Grande do Norte - UFRN in Revista de Geociências do Nordeste
- Vol. 12 (1) , 340-365
Modelagem implícita e geoestatística aplicadas à construção de um modelo geotécnico para uma mina a céu aberto no Brasil
O conceito de modelo geotécnico em minas a céu aberto baseia-se em um produto definido por domínios ou unidades com comportamento geotécnico similar. O principal objetivo é apoiar os processos de design de taludes. A construção de um modelo geotécnico deve integrar quatro outros modelos: geológico, estrutural, geomecânico e hidrogeológico. Este artigo visa apresentar uma metodologia para a construção desses modelos utilizando geoestatística e modelagem implícita, e posteriormente, a integração desses modelos para criar um modelo de bloco geotécnico unificado. Este modelo contém informações sobre índice de resistência, grau de fraturamento, rugosidade, Índice de Resistência Geológica, anisotropia e condutividade hidráulica. A metodologia desenvolvida provou ser eficaz na integração dos modelos temáticos produzidos. A abordagem geoestatística reduziu a subjetividade na construção dos modelos, tornando este processo mais quantitativo. Este foi um dos principais diferenciais desta metodologia, pois permite a classificação da confiabilidade das informações geotécnicas atribuídas a cada bloco.
- 5 January 2026
- Published by Universidade Federal do Rio Grande do Norte - UFRN in Revista de Geociências do Nordeste
- Vol. 12 (1) , 340-365
Open AccessModelagem implícita e geoestatística aplicadas à construção de um modelo geotécnico para uma mina a céu aberto no Brasil
O conceito de modelo geotécnico em minas a céu aberto baseia-se em um produto definido por domínios ou unidades com comportamento geotécnico similar. O principal objetivo é apoiar os processos de design de taludes. A construção de um modelo geotécnico deve integrar quatro outros modelos: geológico, estrutural, geomecânico e hidrogeológico. Este artigo visa apresentar uma metodologia para a construção desses modelos utilizando geoestatística e modelagem implícita, e posteriormente, a integração desses modelos para criar um modelo de bloco geotécnico unificado. Este modelo contém informações sobre índice de resistência, grau de fraturamento, rugosidade, Índice de Resistência Geológica, anisotropia e condutividade hidráulica. A metodologia desenvolvida provou ser eficaz na integração dos modelos temáticos produzidos. A abordagem geoestatística reduziu a subjetividade na construção dos modelos, tornando este processo mais quantitativo. Este foi um dos principais diferenciais desta metodologia, pois permite a classificação da confiabilidade das informações geotécnicas atribuídas a cada bloco.
- Open Access
- 1 January 2026
- Published by FapUNIFESP (SciELO) in Anais da Academia Brasileira de Ciências
- Vol. 98 (2) , e20241434
A new approach to rock mass classification using machine learning algorithms
This paper proposes a new model for rock mass classification, using machine learning techniques. The variables used are often associated with well-known geomechanical classifications. Factor analysis permitted reducing system dimensionality, selecting only the significant variables related to the rock mass quality. The proposed model used the k-medoid clustering technique to create the geomechanical classes, specifically the partitioning around medoids algorithm, unsupervised learning. The results indicated the formation of seven groups, labeled through the interpretation of the variable values in each group. Subsequently, a decision tree was applied to obtain the geomechanical classes defined by k-medoids technique; aiming to present an easy way for new users to apply the proposed rock mass classification. The proposed model is a new approach to deal with rock mass classification problems, decreasing subjectivity, increasing parameter selectivity, reducing the dimensionality of geotechnical databases and optimizing the application of classification systems.
- 1 January 2026
- Published by FapUNIFESP (SciELO) in Anais da Academia Brasileira de Ciências
- Vol. 98 (2) , e20241434
Open AccessA new approach to rock mass classification using machine learning algorithms
This paper proposes a new model for rock mass classification, using machine learning techniques. The variables used are often associated with well-known geomechanical classifications. Factor analysis permitted reducing system dimensionality, selecting only the significant variables related to the rock mass quality. The proposed model used the k-medoid clustering technique to create the geomechanical classes, specifically the partitioning around medoids algorithm, unsupervised learning. The results indicated the formation of seven groups, labeled through the interpretation of the variable values in each group. Subsequently, a decision tree was applied to obtain the geomechanical classes defined by k-medoids technique; aiming to present an easy way for new users to apply the proposed rock mass classification. The proposed model is a new approach to deal with rock mass classification problems, decreasing subjectivity, increasing parameter selectivity, reducing the dimensionality of geotechnical databases and optimizing the application of classification systems.
- Open Access
- 1 January 2025
- Published by FapUNIFESP (SciELO) in Anais da Academia Brasileira de Ciências
- Vol. 97 (1) , e20240426
A new approach to dilution prediction of underground mine gold using computing techniques
Controlling ore dilution in underground mining is challenging. In this study, data from a Brazilian gold mine were analyzed, covering 70 chambers and 26 variables. Six key variables were identified through decision tree analysis, forming the basis of a predictive model using advanced soft computing techniques. The constructed Random Forest model (RF-A) significantly outperformed two predictive equations from the literature, achieving an R² of 0.9161 compared to 0.3009 and 0.1597 from the literary equations. Validation of RF-A with random subsampling resulted in a marginal decrease in the R² value to 0.3060, suggesting a nonlinear correlation between mining variables and dilution, highlighting the inadequacy of linear analysis methods. By dividing the dataset into three subsets representing different mineral bodies, three new Random Forest models (RF-CV, RF-CB, and RF-LJ) were created, with R² values of 0.5465, 0.5295, and 0.4525, respectively. These results underscore the need to tailor models to specific geological contexts and demonstrate the potential of machine learning techniques in predicting dilution in complex underground mining scenarios.
- 1 January 2025
- Published by FapUNIFESP (SciELO) in Anais da Academia Brasileira de Ciências
- Vol. 97 (1) , e20240426
Open AccessA new approach to dilution prediction of underground mine gold using computing techniques
Controlling ore dilution in underground mining is challenging. In this study, data from a Brazilian gold mine were analyzed, covering 70 chambers and 26 variables. Six key variables were identified through decision tree analysis, forming the basis of a predictive model using advanced soft computing techniques. The constructed Random Forest model (RF-A) significantly outperformed two predictive equations from the literature, achieving an R² of 0.9161 compared to 0.3009 and 0.1597 from the literary equations. Validation of RF-A with random subsampling resulted in a marginal decrease in the R² value to 0.3060, suggesting a nonlinear correlation between mining variables and dilution, highlighting the inadequacy of linear analysis methods. By dividing the dataset into three subsets representing different mineral bodies, three new Random Forest models (RF-CV, RF-CB, and RF-LJ) were created, with R² values of 0.5465, 0.5295, and 0.4525, respectively. These results underscore the need to tailor models to specific geological contexts and demonstrate the potential of machine learning techniques in predicting dilution in complex underground mining scenarios.
- Open Access
- 27 December 2023
- Published by Instituto Federal de Educacao, Ciencia e Tecnologia do Rio Grande do Norte (IFRN) in HOLOS
- Vol. 8 (39)
MAPAS AUTO-ORGANIZÁVEIS APLICADOS AO DESAGRUPAMENTO EM AMOSTRAGEM PREFERENCIAL
Os processos de amostragem na exploração mineral muitas vezes resultam em áreas preferencialmente amostradas, com a formação de agrupamentos, que podem surgir devido a alguns fatores, tais como condições de acessibilidade, valores de atributos e a estratégia de amostragem. Os agrupamentos afetam a inferência estatística da área. O objetivo deste artigo é propor uma nova abordagem para métodos de desagrupamento usando as redes de Kohonen, Self-Organizing Maps (SOM). As SOMs é um tipo de rede neural artificial usada para classificação não supervisionada. A metodologia atribui a cada amostra um peso para calcular a média desagrupada. A atribuição de peso para cada amostra em uma área é inversamente proporcional à área densamente amostrada. A média desagrupada é dada pela soma da multiplicação do peso com o valor do atributo de cada amostra. Portanto, a lógica de atribuição de pesos é semelhante ao método Cell Declustering, porém as SOMs identificam as áreas com margens não lineares, ao contrário do método Cell Declustering. Um estudo de caso é apresentado, usando o conjunto de dados de Walker Lake. A presente pesquisa não pretende substituir os métodos clássicos de desagrupamento, mas sim apresentar uma nova abordagem para um problema rotineiro na avaliação de reservas. Embora a matemática da técnica aplicada seja de fato complexa, os resultados podem ser promissores.
- 27 December 2023
- Published by Instituto Federal de Educacao, Ciencia e Tecnologia do Rio Grande do Norte (IFRN) in HOLOS
- Vol. 8 (39)
Open AccessMAPAS AUTO-ORGANIZÁVEIS APLICADOS AO DESAGRUPAMENTO EM AMOSTRAGEM PREFERENCIAL
Os processos de amostragem na exploração mineral muitas vezes resultam em áreas preferencialmente amostradas, com a formação de agrupamentos, que podem surgir devido a alguns fatores, tais como condições de acessibilidade, valores de atributos e a estratégia de amostragem. Os agrupamentos afetam a inferência estatística da área. O objetivo deste artigo é propor uma nova abordagem para métodos de desagrupamento usando as redes de Kohonen, Self-Organizing Maps (SOM). As SOMs é um tipo de rede neural artificial usada para classificação não supervisionada. A metodologia atribui a cada amostra um peso para calcular a média desagrupada. A atribuição de peso para cada amostra em uma área é inversamente proporcional à área densamente amostrada. A média desagrupada é dada pela soma da multiplicação do peso com o valor do atributo de cada amostra. Portanto, a lógica de atribuição de pesos é semelhante ao método Cell Declustering, porém as SOMs identificam as áreas com margens não lineares, ao contrário do método Cell Declustering. Um estudo de caso é apresentado, usando o conjunto de dados de Walker Lake. A presente pesquisa não pretende substituir os métodos clássicos de desagrupamento, mas sim apresentar uma nova abordagem para um problema rotineiro na avaliação de reservas. Embora a matemática da técnica aplicada seja de fato complexa, os resultados podem ser promissores.
- Open Access
- 1 March 2023
- Published by FapUNIFESP (SciELO) in REM - International Engineering Journal
- Vol. 76 (1) , 71-78
Assessment of mine slopes stability conditions using a decision tree approach
- 1 March 2023
- Published by FapUNIFESP (SciELO) in REM - International Engineering Journal
- Vol. 76 (1) , 71-78
Open AccessAssessment of mine slopes stability conditions using a decision tree approach
- Open Access
- 17 October 2022
- Published by Research, Society and Development in Research, Society and Development
- Vol. 11 (14)
Identificação de padrões morfoestruturais utilizando Clustering Large Applications, um estudo de caso no Quadrilátero Ferrífero
Dentre as etapas de um projeto de mineração destaca-se a pesquisa mineral, com objetivos de identificar, estudar e avaliar os depósitos minerais. Nesta etapa específica ocorre a transformação dos recursos minerais inferidos, em indicados e por fim medidos, e caso seja viável sua explotação, em reservas minerais prováveis e/ou provadas. A descoberta destas reservas é marco impactante para o desenvolvimento industrial, tecnológico e econômico de uma sociedade. Este artigo tem como objetivo principal apresentar a utilização de uma técnica de machine learning para identificação de estruturas de particular interesse geológico, a partir de imagens de satélite. A técnica aplicada foi o Clustering Large Applications (CLARA) que é um algoritmo não-supervisionado para agrupamento de dados, com alta performance em banco de dados massivos. A área utilizada como estudo de caso foi o Quadrilátero Ferrífero, uma das maiores províncias minerais do planeta, localizada no estado de Minas Gerais, Brasil. Os resultados do modelo CLARA permitiram delinear todas as feições que formam o Quadrilátero Ferrífero. Neste contexto acredita-se que esta possa ser uma boa ferramenta para seleção de alvos exploratórios reduzindo incerteza e risco aos investidores. O que propicia não somente a atração de novas empresas para pesquisa mineral, além da ampliação das reservas dos recursos minerais brasileiros.
- 17 October 2022
- Published by Research, Society and Development in Research, Society and Development
- Vol. 11 (14)
Open AccessIdentificação de padrões morfoestruturais utilizando Clustering Large Applications, um estudo de caso no Quadrilátero Ferrífero
Dentre as etapas de um projeto de mineração destaca-se a pesquisa mineral, com objetivos de identificar, estudar e avaliar os depósitos minerais. Nesta etapa específica ocorre a transformação dos recursos minerais inferidos, em indicados e por fim medidos, e caso seja viável sua explotação, em reservas minerais prováveis e/ou provadas. A descoberta destas reservas é marco impactante para o desenvolvimento industrial, tecnológico e econômico de uma sociedade. Este artigo tem como objetivo principal apresentar a utilização de uma técnica de machine learning para identificação de estruturas de particular interesse geológico, a partir de imagens de satélite. A técnica aplicada foi o Clustering Large Applications (CLARA) que é um algoritmo não-supervisionado para agrupamento de dados, com alta performance em banco de dados massivos. A área utilizada como estudo de caso foi o Quadrilátero Ferrífero, uma das maiores províncias minerais do planeta, localizada no estado de Minas Gerais, Brasil. Os resultados do modelo CLARA permitiram delinear todas as feições que formam o Quadrilátero Ferrífero. Neste contexto acredita-se que esta possa ser uma boa ferramenta para seleção de alvos exploratórios reduzindo incerteza e risco aos investidores. O que propicia não somente a atração de novas empresas para pesquisa mineral, além da ampliação das reservas dos recursos minerais brasileiros.
- 26 October 2021
- Published by Springer Nature in Neural Computing and Applications
- Vol. 34 (6) , 4633-4642
Evaluation of machine learning methods for rock mass classification
- 26 October 2021
- Published by Springer Nature in Neural Computing and Applications
- Vol. 34 (6) , 4633-4642
Evaluation of machine learning methods for rock mass classification