This study employed machine learning to investigate the mechanical behavior of one-part geopolymer (OPG)-stabilized soil subjected to acid erosion. Based on the unconfined compressive strength (UCS) data of acid-eroded OPG-stabilized soil, eight machine learning models, namely, Adaptive Boosting (AdaBoost), Decision Tree (DT), Extra Trees (ET), Gradient Boosting (GB), Light Gradient Boosting Machine (LightGBM), Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost), along with hyper-parameter optimization by Genetic Algorithm (GA), were used to predict the degradation of the UCS of OPG-stabilized soils under different durations of acid erosion. The results showed that GA-SVM (R² = 0.9960, MAE = 0.0289) and GA-XGBoost (R² = 0.9961, MAE = 0.0282) achieved the highest prediction accuracy. SHAP analysis further revealed that solution pH was the dominant factor influencing UCS, followed by the FA/GGBFS ratio, acid-erosion duration, and finally, acid type. The 2D PDP combined with SEM images showed that the microstructure of samples eroded by HNO₃ was marginally denser than that of samples eroded by H₂SO₄, yielding a slightly higher UCS. At an FA/GGBFS ratio of 0.25, abundant silica and hydration products formed a dense matrix and markedly improved acid resistance. Further increases in FA content reduced hydration products and caused a sharp drop in UCS. Extending the erosion period from 0 to 120 days and decreasing the pH from 4 to 2 enlarged the pore network and diminished hydration products, resulting in the greatest UCS reduction. The results of the study provide a new idea for applying the ML model in geoengineering to predict the UCS performance of geopolymer-stabilized soils under acidic erosion.

Geopolymers, as an eco-friendly alternative construction material to ordinary Portland cement (OPC), exhibit superior performance in soil stabilization. However, their inherent brittleness limits engineering applications. To address this, polyethylene terephthalate (PET) fibers can be incorporated into a one-part geopolymer (OPG) binder to enhance ductility while promoting plastic waste recycling. However, the evaluation of ductile behavior of OPG-stabilized soil with PET fiber normally demands extensive laboratory and field experiments. Leveraging artificial intelligence, a predictive model can be developed for this purpose. In this study, data were collected from compressive and tensile tests performed on the OPG-stabilized soil with PET fiber. Four deep learning neural network models, namely ANN, BPNN, CNN, and LSTM, were then used to construct prediction models. The input parameters in the model included the fly ash (FA) dosage, dosage and length of the PET fiber, and the Curing Time. Results revealed that the LSTM model had the best performance in predicting the three ductile properties (i.e., the compressive strength index [UCS], strain energy index [CSE], and tensile strength index [TES]). The SHAP and 2D-PDP methods were further used to verify the rationality of the LSTM model. It is found that the Curing Time was the most important factor for the strength and ductile behavior. The appropriate addition of PET fiber of a certain length had a positive impact on the ductility index. Thus, for the OPG-stabilized soil, the optimal dosage and length of PET fiber were found to be 1.5% and 9 mm, respectively. Additionally, there was a synergistic effect between FA and PET on the ductility metric. This research provides theoretical support for the application of geopolymer and PET fiber in enhancing the ductility of the stabilized soil.

Concrete mix design increasingly requires rapid screening of mixture proportions against mechanical and resource-efficiency targets. This study develops an explainable, cement-reduction-oriented computational screening framework based on the UCI Concrete Compressive Strength dataset. The dataset records Portland cement, fly ash, blast furnace slag, aggregates, water, superplasticizer, curing age, and compressive strength, but does not report clinker factor, material-specific emission factors, or durability performance. Cement dosage is therefore minimized only as a surrogate objective; the study does not claim quantified embodied-carbon optimization. A LightGBM surrogate was trained using raw mixture variables and domain-informed ratios. On the held-out test set, the model achieved an $R^2$ of 0.946 and a mean absolute error of 2.639 MPa. Shapley Additive Explanations were used to examine the statistical influence of mixture variables, with the water-to-binder ratio emerging as the dominant predictor. Constraint-filtered Monte Carlo sampling and non-dominated sorting were then used to screen Pareto-efficient binder allocations. For a 28-day target of 45 MPa, repeated searches identified candidates with a mean Portland cement dosage of $164.4 {\mathrm{kg/m}}^3$, 44.2% below the mean of empirical mixtures in the same strength band. The evaluated numerical modules were embedded in a Streamlit prototype in which a DeepSeek large language model performs only intent parsing and report generation. The main contribution is this tool-augmented separation of language interaction from deterministic engineering computation. The resulting mixtures remain computational candidates and should next be validated experimentally and assessed using material-specific life-cycle carbon and durability data.

Predicting the unconfined compressive strength (UCS) of alkali-activated stabilized soft soil under coupled acid attack and wetting–drying cycles requires evaluation across correlated laboratory observations. This study analyzed 689 records from 53 complete mixtures, with seven material and environmental inputs, using four machine-learning models (DT, RF, SVM, and XGBoost) and four neural models (PINN, BPNN, CNN, and LSTM). Generalization to previously unseen mixtures was evaluated by five-fold outer GroupKFold with three-fold mixture-grouped tuning. Each original record contributed one out-of-fold prediction. The pooled grouped-CV $R^2$ values ranged from 0.5043 to 0.7605, with RMSEs of 0.3579–0.5148 MPa. XGBoost attained the highest point estimate ($R^2=0.7605$; RMSE = 0.3579 MPa), followed by SVM ($R^2=0.7362$; RMSE = 0.3756 MPa). PINN and its architecture-matched BPNN produced identical grouped-CV predictions and no demonstrated incremental benefit. A prospective penalty-activation diagnostic confirmed that soft penalties can affect training trajectories but did not reconstruct the archived model-selection runs. A separately tuned, five-step, structure-matched ablation found no additional predictive gain from preceding design covariates in CNN and LSTM. Historical fixed-configuration random-CV comparisons were removed because the original hyperparameter-selection provenance could not be verified. SHAP and partial dependence analyses describe model-learned associations, and SEM offers qualitative morphological context. These models provide preliminary screening aids within the tested materials and factor levels, not engineering design tools; independent batch and field validation remains necessary.