Open Access

ARTICLE

Explainable GAN-Augmented MLP for Soil Resilient Modulus Prediction

Huiguo Wu^1,2,3, Yuedong Wu^1,2,3,*, Jian Liu^2,3

1 College of Artificial Intelligence and Automation, Hohai University, Changzhou, China
2 College of Civil and Transportation Engineering, Hohai University, Nanjing, China
3 Key Laboratory of Ministry of Education for Geomechanics and Embankment Engineering, Hohai University, Nanjing, China

* Corresponding Author: Yuedong Wu. Email:

(This article belongs to the Special Issue: Advances in Artificial Intelligence for Geotechnical Engineering)

Computer Modeling in Engineering & Sciences 2026, 148(1), 27 https://doi.org/10.32604/cmes.2026.083977

Received 14 April 2026; Accepted 22 June 2026; Issue published 27 July 2026

Abstract

The resilient modulus (M_R) is a key mechanical parameter in geotechnical engineering, but conventional laboratory measurement is time-consuming and labor-intensive. Deep learning models provide an alternative for predicting M_R using easily obtainable soil properties, yet their performance is often limited by the small size of available datasets. To address this limitation, this study develops an interpretable data-enhanced deep learning framework for M_R prediction. In the proposed framework, a multilayer perceptron (MLP) is adopted as the base prediction model, a generative adversarial network (GAN) is used to generate synthetic samples from the limited training data, Optuna is employed for hyperparameter optimization, and Shapley Additive Explanations (SHAP) are adopted to interpret the trained model. Random Forest (RF) and extreme gradient boosting (XGBoost) models optimized by Optuna are also introduced for performance comparison. The results show that the GAN-generated samples are generally consistent with the original data in terms of distribution characteristics and correlation patterns. Compared with the Optuna-MLP model without GAN-based data augmentation, the proposed GAN-Optuna-MLP model improves the testing R² from 0.87 to 0.93, reduces the mean absolute error (MAE) from 3.62 to 2.72 MPa, and decreases the mean absolute percentage error (MAPE) from 7.61% to 5.70%. The proposed model also achieves comparable performance to the optimized RF and XGBoost models, with slightly lower error-based indicators. In addition, SHAP analysis identifies cone tip resistance as the most influential variable for M_R prediction. These findings suggest that the integration of data augmentation, automated optimization, and interpretability analysis offers a practical solution for small-sample soil property prediction.

---

Keywords

---