A multifidelity neural network (MFNN) for constitutive modeling of complex soil behaviors

The development and calibration of soil models under the framework of plasticity is notoriously challenging given the prismatic features in soil's shear behaviors. Data-driven deep neural networks (DNNs) offer an alternative approach to this formidable task. However, classical DNN models struggle to accurately capture soil mechanical responses using limited training data. To address this issue, a unified multifidelity neural network (MFNN) is proposed to leverage the accuracy of high-fidelity experimental datasets and the abundance of low-fidelity datasets synthesized using the modified Cam-clay (MCC) model. The MFNN model can automatically learn the correlation between the low-fidelity and high-fidelity datasets, and has been applied to predicting the mechanical responses of over-consolidated and anisotropically consolidated clays, cement-treated clays, silts sheared under varying temperatures, and suction-controlled unsaturated soils. The results demonstrate significantly improved prediction capabilities of MFNN compared to purely data-driven DNNs. The proposed MFNN framework holds encouraging promise to revolutionize constitutive modeling of complex soil behaviors.