Investigating the Potential of Machine Learning Methods to Predict Soil Variables for Dike's Macro Stability Analysis
Master thesis research, not a Qynzoo client project. Our founder Mostafa did this research for his master thesis at the University of Twente, using data from Fugro. Fugro was not a Qynzoo client. We show it because this research is relevant to the services we offer, and every solution we build is made from scratch for your business. Read the full thesis in the University of Twente repository →
Random Forest, Neural Networks and Auto-Encoders were used to predict dike clay's dry unit weight and friction angle, giving designers a data-driven estimate before slow and expensive lab tests are run.
MASTER THESIS · MACHINE LEARNING
Soil variables estimated before the lab tests
GEOTECHNICAL DATA
EARLY INPUTS
MACHINE LEARNING
INSIGHTS + REPORT
Fugro database, Dutch dikes
Atterberg limits
Clay / sand / silt
Water content
Random Forest
Neural Network
Auto-encoder
Trained across 4 scenarios
Dry unit weight: predicted
Friction angle: data too limited
ACADEMIC REPORT
Random Forest gave the strongest predictions, especially for dry unit weight.
Designers get a data-driven estimate before slow, expensive lab tests.
Animation: geotechnical data from Fugro's database and early inputs such as Atterberg limits, soil mix and water content go into Random Forest, Neural Network and Auto-encoder models. Dry unit weight is predicted, the friction angle data turned out too limited, and the findings are written up in an academic report.
MASTER THESIS · MACHINE LEARNING
Soil variables estimated before the lab tests
GEOTECHNICAL DATA
Fugro database, Dutch dikes
EARLY INPUTS
Atterberg limits
Clay / sand / silt
Water content
MACHINE LEARNING
Random Forest
Neural Network
Auto-encoder
Trained across 4 scenarios
INSIGHTS + REPORT
Dry unit weight: predicted
Friction angle: data too limited
ACADEMIC REPORT
Random Forest gave the strongest predictions, especially for dry unit weight.
Designers get a data-driven estimate before slow, expensive lab tests.
Animation: geotechnical data from Fugro's database and early inputs such as Atterberg limits, soil mix and water content go into Random Forest, Neural Network and Auto-encoder models. Dry unit weight is predicted, the friction angle data turned out too limited, and the findings are written up in an academic report.
The Problem
Designing a dike's macro-stability requires knowing the clay's dry unit weight and friction angle. Today, engineers get these from triaxial and compaction tests, which are expensive, slow and not available during the design phase. Until the tests are run, designers fall back on manual estimates based on clay/sand ratio and assumed consistency (weak, moderate, strong). This project asked: can machine learning make that design-phase estimate more precise, using only the data an engineer already has on hand?
The Approach
Using Dutch geotechnical data from Fugro's database, Random Forest and Neural Network models were trained across four scenarios reflecting different levels of design-phase knowledge: Atterberg limits, clay/sand/silt content, NEN 5014 triangular classification, and plasticity diagram category. Auto-encoders were tested for feature extraction. The goal was to find which combination of inputs and models best predicted dry unit weight and friction angle before any lab test is run.
Neural NetworksRandom ForestAuto-EncodersPython
Results
Random Forest outperformed Neural Networks across scenarios, with the Atterberg-limits scenario producing the strongest predictions, particularly for dry unit weight. Combining this with water content estimates at three consistency indices (0.60 / 0.75 / 0.85) allowed unit weight to be predicted with low uncertainty, directly mirroring how engineers reason through the estimate today, but with better accuracy.
Friction angle could not be reliably predicted. This was not a modelling failure but a data one: the available triaxial dataset was too small and inconsistent in quality to build any meaningful correlation. The models simply had nothing to learn from. The practical conclusion still stands: unit weight estimation can realistically skip compaction and triaxial testing in the design phase, cutting cost and time without sacrificing accuracy.