In this paper, we combined the traditional ISTA approach with deep learning, transferred the application domain from image inverse problem to ECG inverse problem because of its excellent performance to propose a new ISTA-net method to solve the inverse problem of ECG. The proposed method has quick convergence speed, steady performance and good extend capability. We created the ISTA-net algorithm model based on the ECG physical model, and transformed it into a neural network structure to achieve automatic parameter selection and more accurate result. We use a real human model on the ECGsim software to obtain the simulated potentials on the cardiac surface and body surface to train our model. The result of our experiment confirms that the proposed method has a better performance in reconstructing epicardial surface potentials distribution compared with the common regularization method like TTLS, Tikhonov, TSVD method.