Work

Testing the applicability of machine learning strength models for stabilized dredged sludge and slurry to new materials

Year: 2026
Type: preprint
Abstract: Literature-trained machine learning models support mixture design for stabilized dredged sludge and slurry, but random record splits test them on studies already in training rather than on new materials. We combined leave-one-source-out (LOSO) validation, a source-coverage comparison, within-material water-content pairs and comparisons of calibrated and local models. Models trained on 399 unconfined compressive strength records from 10 sources were tested on four new materials and replicated with alkali-activated concrete (AAC) data. The random forest’s LOSO mean absolute error (MAE) was 3.74 times its record-random-split median. Replacing the predicted source’s training records with an equal number from other sources raised the error 2.57-fold (1.77-fold for AAC). A random forest using stabilizer dosage, age and water content had lower LOSO error than one using all inputs but predicted zero or less accurate within-material strength changes for some materials. With 10 local records, the calibrated random forest had a lower median MAE than raw-scale local ridge regression for two of three materials, whereas Gaussian process regression, best under LOSO validation, did so for none. For a new material, error should be assessed by LOSO validation, water-content effects by within-material pairs and calibration against local models on the same records.
Language: English
Open Access status: green
OpenAlex ID: W7235395814
Cites: 39
Cited by:
Related to: 10