With artificial intelligence (AI), scientists and engineers are faced with a familiar question: The tools are impressive on their own, but how do we use them to provide effective, reproducible results? Our recent work for a client to develop a groundwater model that was as accurate as traditional modeling but with fewer data points and lower cost helps illustrate the way forward.

For this project, we focused on machine learning (ML), which is a subset of AI. ML uses algorithms and statistical models to analyze data for patterns, draw inferences from those patterns, and learn from the patterns without being issued explicit instructions. In this instance, we needed to demonstrate to regulators that impacted groundwater was discharging to a nearby stream and not flowing underneath the stream.  

Traditional modeling to meet our objective required obtaining a significant amount of information through complex field work and gathering historical information, LIDAR data, geological data, etc. The information needs for our ML model were significantly less and limited to weather information from a nearby weather station, river elevation data, and groundwater levels. Our AI-assisted model used Python, a common programming language popular in the ML community, and an open source “package” named Pastas developed for groundwater scientists and engineers to analyze hydrogeological time series. Pastas uses a transfer function noise model to show how the groundwater system will respond to a  stressor (e.g., precipitation) while also incorporating random noise on the output to better reflect the complexity of a hydrogeological system. 

To test the model, we used it to predict future groundwater elevations which were compared to actual measured values over time. Using the Normalized Root Mean Square Error (NRMSE), which compares the predicted values against the observed values while normalizing them by the standard deviation of the observed data, we found that our ML model’s values came within 2% of real-world observations.

Our approach has the potential to be used in a wide range of groundwater scenarios, from estimating snowmelt effects to controlling groundwater in tunneling operations. Core to going forward, however, will be ensuring that the model reflects the real-world data, and that the data reflects the risks: In other words, trust, but verify. 

Scientists and clients are understandably concerned about how to use these tools and where they fit into our work. By developing ML models that can be checked against real-world results, and carefully choosing what data a model is built on, clients and contractors alike can make better-informed decisions, while saving time and money.

The Author:

Matthew J. Gozdor is a Quantitative Hydrogeologist and a Senior Technical Specialist at GZA GeoEnvironmental, Inc. With 25 years of experience, his focus is on groundwater flow modeling, fate & transport modeling, aquifer testing design and evaluation, and hydrologic evaluations. www.gza.com, matthew.gozdor@gza.com