Newton Tunnel: Real-Time Ground Monitoring

The Story

BC's water infrastructure serves thousands of people. In Newton, a 1.5-meter tunnel had to bore 150 meters underground:directly beneath two live water reservoirs. One wrong prediction would trigger catastrophe: the tunnel collapses, water supply shuts down, or drilling delays cost $50,000 per day.

The drilling crew had geological surveys, but predicting ground conditions 10 meters ahead of the drill bit required understanding what they were about to hit before they hit it. They asked WSP: Can we predict what's coming?

I built a system that did exactly that.

The Problem

• Tunnel bore beneath two live water reservoirs:zero margin for error
• Ground conditions unknown and variable:past surveys can't predict present conditions
• Drilling decisions made with limited look-ahead data
• One wrong call = $50K/day in downtime or water supply loss
• Crew needed daily predictions for shift planning

The Solution

I built a predictive system that combined geology, historical drilling data, and real-time sensor readings to forecast ground conditions ahead of the drill.

Data Sources

• Geological surveys: 12 years of drilling data from the region:bore logs, rock types, water tables
• Borehole data: Pre-construction sampling showed soil composition at specific depths
• Real-time sensor data: Penetration rate, torque, vibration:signals of what the drill was encountering

The Model

Integrated these data sources into a probabilistic ground model. For every meter the drill advanced, the model recalibrated using the latest sensor readings. The output: a probability distribution of what soil conditions were coming (sand, clay, rock, water), updated daily.

Delivery

Daily reports showed the crew: (1) What they drilled through today, (2) What to expect in the next 10 meters, (3) Risk zones and mitigation strategies. Crew could plan drilling pace, equipment, and personnel accordingly.

The Impact

The tunnel finished on schedule. Zero delays. Zero water supply incidents. The crew's confidence in the predictions meant they could move faster without second-guessing.

The Build

My role: Project Engineer leading data integration. Worked with geotechnical specialists, drilling engineers, and instrumentation teams. Built the prediction pipeline and delivered daily reports.

Tech Stack

Workflow

• Daily data collection: Sensor readings from the TBM (tunnel boring machine) uploaded to a database each shift
• Model update: Python pipeline ingested new data, recalibrated the model with latest conditions
• Prediction generation: Model predicted soil conditions for the next 10 meters and probability of water zones
• Report delivery: Visual summary (charts + maps) delivered each morning to the site engineer

Key Decisions

• Focused on probability, not certainty. The crew understood "60% chance of clay" better than a point prediction.
• Daily updates, not real-time. The model was only as good as the geological knowledge:updating too often would create false confidence.
• Simple deliverable. A one-page report with maps and predictions. No jargon for the site team.

Learnings

• Domain knowledge beats model complexity: A simple statistical model + 12 years of drilling history beat an overfitted ML model. Geology has patterns:respect them.
• Probability is clearer than precision: Crews care about risk ranges, not decimal-point accuracy. "High risk of water" is actionable; "68.47% saturation" is noise.
• Sensor data needs interpretation: TBM readings tell you what you're hitting, but you have to translate that into actionable drilling decisions. Raw data doesn't help.
• Build for the field, not the office: Deliverables need to work for tired crews on site, not just look good in reports. PDF with maps beats a dashboard.
• Under-promise on impact: I said the system would "inform decisions." I didn't claim it would eliminate risk. That trust is what made the crew listen.