Typical estimation of petrophysical properties in a reservoir volume is carried out by seismic inversion with integration of rock physics models. While this provides good solutions to the original objectives, it is strongly dependent on data-fitting algorithms to create an empirical model during well analyses. To overcome this limitation, a cross-shape deep Boltzmann machine structure is designed by arranging four different restricted Boltzmann machines located at the vertices interconnected via a hidden neuron layer at the center. Four different input data types, including a seismic amplitude and three petrophysical properties logs (porosity, water saturation, and shale volumetric), are used to train the network. Ultimately, six different nonlinear relationships between the inputs are simultaneously captured during the learning process, among which the three connections between target petrophysical logs and seismic amplitudes are used to predict property values. Unlike many conventional deep learning systems, which expect large data for all what-if scenarios into training, this multilayer probabilistic network does not involve any backpropagation, but only requires the anticipated upper and lower bounds of the data and labels for random initializations of Markov chains to reach an equilibrium stage. This makes the algorithm an excellent candidate for solving problems with limited data sets. This approach is robust enough to account for uncertainties while still retaining the generalization outside of the training information. A two-dimensional field data set with limited well coverage is used to demonstrate the capability of this algorithm with accurate reconstruction of geologically plausible petrophysical property sections.