This paper proposes a procedure based on machine learning for the inverse analysis of thermophysical parameters of Malan loess using finite element method as a trainer. A novel neural network is established to identify the thermophysical parameters relying on nodal temperatures of the loess block. The ill-posed problem can be solved stably and accurately using the present neural network. Besides, this inverse network can determine multiply physical parameters at the same time. In this work, a finite element model is built first for the direct transient heat conduction problem of the Malan loess, which is used as the trainer to create the samples for the inverse analysis. Then, the thermophysical parameters of the samples are generated by the Latin hypercube sampling. Based on thermophysical parameter samples, the corresponding nodal temperature samples can be calculated using the proposed finite element model for the inverse analysis. Using these samples, a completely-fully connected neural network called Parm-Net is constructed and trained for this inverse problem. The effects of the learning rate, batch size and number of neurons of Parm-Net are also discussed. Finally, intensive numerical experiments are carried out to demonstrate the accuracy and efficient of Parm-Net. The results of Parm-Net show that it works well with the error less than 4%, even in the nodal temperatures with large noisy.