Applied Random Forest, Neural Networks, and Auto-Encoders to predict dike clay's dry unit weight and friction angle — replacing slow, expensive lab tests with a data-driven design-phase estimate.
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 — both expensive, slow, and unavailable 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?
Using Dutch geotechnical data from Fugro's database, I trained Random Forest and Neural Network models 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.
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 currently reason through the estimate — just with better accuracy.
Friction angle could not be reliably predicted — not due to a modeling 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.