<p>Compressibility of reconstituted clay generally plays a critical role in evaluating the workability of geo-structures. Previous empirical models may show limited accuracy beyond the range of database. Meanwhile, pure data-driven models may not well reproduce mechanical behavior of soils even with sufficient data. Therefore, a multi-fidelity neural network (MFNN) is proposed to capture the compression behavior of reconstituted clay. The synthetic data generated by empirical equations are utilized for training the low-fidelity model (LFM), while the high-fidelity model (HFM) is trained by tests results. This training strategy enables the MFNN to combine the easy accessibility of synthetic data and high accuracy of the test results. Initial water content, liquid limit, and effective vertical stress are adopted as input variables for predicting the void ratio. The LFM provides a baseline for the whole model and the HFM is decomposed into linear and nonlinear parts for a more precise prediction. The parameters (i.e., weights and biases) of sub-networks of the MFNN are updated independently to obtain optimal outputs of each network. The final predictions are further obtained by a weighted optimization of linear and nonlinear networks outputs. Results indicate that the distinctive structure of MFNN can reduce the requirement of the data for model training and maintain the generalization ability of the model simultaneously. Only two sets of experimental data are required for satisfied predictions within a wide range of initial water content.</p>