Estimating treatment effects from observational data through covariate matching remains an active research area in causal inference. Although existing methods may provide accurate results on simulated datasets, knowing how to tune the parameters to accurately estimate treatments in practice can be a challenge, since the ground truth is not known. We provide an explainable hybrid neural network and self-organizing map (SOM) approach, Match2. Using a supervised learning paradigm, our method simultaneously learns a meaningful latent representation with respect to treatment assignment and a nonlinear neighborhood preserving mapping via SOM in the latent space. To select the appropriate latent dimension, we propose a data-driven strategy based on the minimum validation loss for the treatment classification subproblem. Unlike other matching methods, the hybrid SOM-neural network can be used as the basis for visualizing and quantifying the quality of the matches. The user can understand the quality of the matches to provide confidence in the results and detect any potential problems. We design a novel metric to examine the overall quality of matching along with the visualization. We demonstrate strong performance on four benchmark datasets compared to non-neural-network baselines. Integrating a SOM component may potentially benefit other state-of-the-art neural network models for causal effect estimation by gaining interpretability while retaining prediction/estimation accuracy.