Elevating the role of the ground engineer
The changing role of the geotechnical engineer, as artificial intelligence promises new capabilities.
Will the geotechnical engineer’s role look the same in five years’ time, as it has for the last decade? As artificial intelligence (AI) capabilities accelerate at breakneck speed, I’m not so sure.
Geotechnical engineers onsite today still face a daily struggle with managing sensor data, inspections, models and reports. Critical data can be missed simply because there is just too much to process.
Just a few years ago, the AI tools we have today would have seemed like something out of a science-fiction film. The pace of advancement has been extraordinary. Large language models (LLMs) have shown that AI can interpret complex instructions and chain together multiple tools to accomplish goals. The next generation promises even stronger reasoning capabilities.
Additionally, new frameworks are making it easier to integrate AI with existing software. The Model Context Protocol (MCP) allows LLMs to interface directly with complex tools, such as 3D modelling software. These tools are still in their early stages, but how soon will the limiting factor no longer be the technology itself, but how creatively and responsibly we apply it to our domain?
A more realistic near-term development might involve connecting these specialised tools and functions under a single orchestration layer. Instead of juggling multiple applications or data sources, you would interact with a single AI that can use all of them as needed. Here’s how it might work in an open-pit setting:
- Hazard detected – the AI continuously monitors real-time sensor data and spots an unusual deformation pattern in the north wall.
- Context analysis – the system pulls recent geotechnical mapping, inspection reports and historical stability analyses for that sector. It identifies that recent mapping showed weaker geological units not captured in the original model, and an inspection notes new tension cracks forming along the crest.
- Visual core inspection – the AI retrieves core photographs from the drillhole located near the affected area and uses computer vision to assess features. It cross-checks these observations against historical logging data and prior geotechnical reports. Any major discrepancies, such as visually weak zones not reflected in logs or models, are flagged for review by the engineer.
- As-built performance – it automatically checks if the affected area deviated from design specifications. The AI compares survey data with the original design parameters, for example, asking whether catch benches are narrower than planned?
- Stability model update – the AI updates rock properties and pore pressure conditions using the most recent data, then runs a revised stability analysis based on the current pit geometry.
- Output – the engineer receives a report highlighting
The factors impacting stability, such as steeper geometry, weaker materials and rising pore pressures, as well as the model factor of safety. Based on this, the engineer can then determine the appropriate course of action.
Throughout, the system keeps the engineer informed at each stage, explaining where it is pulling data from, what it is analysing and why, ensuring the reasoning can be validated and steering the system as needed. This maintains engineering oversight while dramatically accelerating the analysis workflow.
The key to enabling this level of orchestration is that the AI system has access to bespoke context, functions, tools and domain-specific applications. Each of the specialised sub-agents works in parallel or sequentially to handle each component of the analysis.
If this kind of geotechnical AI assistant can be developed, it could dramatically improve efficiency. These tools will not replace human expertise, but they could change the type of expertise required. We will spend less time wrestling with data and more time applying our expertise where it matters most.
Engineering judgement will be essential when assessing the AI outputs and handling edge cases. Engineers could focus on interpretation and decision-making rather than drowning in data processing and model preparation. There’s also potential to enhance safety, as detailed analysis could be carried out quickly within the operational window to inform actions on the ground.
The end goal is simply to let AI handle the drudgery, so we can focus on solving real-world problems. That said, full autonomy is still a reach. LLMs still need firm guardrails before they can be trusted with critical workflows.
The next few years will be particularly interesting to see if AI systems advance enough to orchestrate true step-by-step analysis workflows. While there are significant challenges around data quality, model validation and ensuring proper oversight, the trajectory seems clear.
