Bridging the Simulation-to-Field Gap in AI-Based Underground Mine Inspection: A Systematic Review of Synthetic Data, Domain Adaptation, Transfer Learning, and Digital Twins
Fatima Umar Zambuk*1, Mustapha Lawal Abdulrahman2, Mahmood Abdulhameed3, Abdulsalam Ya’u Gital4, Saratu Habu Japhet5, Lawal Rukuna Muhammed6
1Dr. Fatima Umar Zambuk, Department of Computer Science, Abubakar Tafawa Balewa University (ATBU), Bauchi, Nigeria.
2Dr. Mustapha Lawal Abdulrahman, National Space Research and Development Agency (NASRDA), Abuja, Nigeria.
3Dr. Mahmood Abdulhameed, Department of Mechatronics Engineering, Abubakar Tafawa Balewa University (ATBU), Bauchi, Nigeria.
4Dr. (Prof.) Abdulsalam Ya’u Gital, Faculty of Engineering, Abubakar Tafawa Balewa University (ATBU), Bauchi, Nigeria.
5Dr. Saratu Habu Japhet, National Space Research and Development Agency (NASRDA), Abuja, Nigeria.
6Dr. Lawal Rukuna Muhammed, Department of Computer Science, Abubakar Tafawa Balewa University (ATBU), Bauchi, Nigeria.
Manuscript received on 21 August 2026 | First Revised Manuscript received on 28 August 2026 | Second Revised Manuscript received on 08 September 2026 | Manuscript Accepted on 15 September 2026 | Manuscript published on 30 September 2026 | PP: 20-24 | Volume-15 Issue-3, September 2026 | Retrieval Number: 100.1/ijrte.C839315030926 | DOI: 10.35940/ijrte.C8393.15030926
Open Access | Editorial and Publishing Policies | Cite | Zenodo | OJS | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open-access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: Artificial intelligence (AI) has advanced rapidly in underground and tunnel inspection, but field deployment continues to be limited by a persistent simulation-to-field reality gap. Real underground data are costly to acquire, difficult to annotate, safety-sensitive, highly site-specific, and often scarce for rare but consequential defects. Simulation, synthetic data, transfer learning, domain adaptation, and digital twins offer complementary ways to address this constraint, yet these research streams are usually reviewed in isolation. This systematic review treats field generalisation as the central analytical problem and synthesises the full pathway from virtual environment construction and synthetic-data generation through cross-domain adaptation, limited-data fine-tuning, digital-twin calibration, and field validation. We used a PRISMA-informed protocol to review literature published from 2015 to 20 August 2026, classifying evidence into three levels: direct underground-mine studies, neardomain tunnel/subterranean studies, and transferable sim-to-real methodology. The synthesis indicates that the most convincing recent results arise not from simulation fidelity alone, but from combining synthetic diversity with target-domain adaptation. Geometry-informed tunnel point-cloud simulation, for example, has shown strong synthetic-only performance and further gains after limited-real-data transfer learning. At the same time, domain-adaptive crack-segmentation studies report substantial performance recovery across materials, structures, and field sites. Mining digital twins are also moving beyond static visualisation toward calibrated virtual-real loops, including mine-tunnel reconstruction, ventilation modelling, rescue-robot navigation, and field-tested teleoperation. Nevertheless, cross-mine validation, explicit measurement of the reality gap, rare-hazard simulation, uncertainty-aware adaptation, and standardised deploymentreadiness criteria remain weak. To address these gaps, this review proposes an Underground Simulation-to-Field Gap Taxonomy and the MineSim2Field framework, an eight-stage closed-loop pathway connecting virtual environment construction, synthetic data, domain diversification, pretraining, adaptation, field calibration, deployment, and twin recalibration. The framework positions simulation as a continuously updated source of evidence rather than a one-off substitute for field data.
Keywords: Underground Mining, Sim-To-Rea, Synthetic Data, Domain Adaptation, Transfer Learning, Digital Twin, Autonomous Inspection, Deep Learning, Field Generalisation, Tunnel Inspection.
Scope of the Article: Computer Science and Engineering’