The efficacy of sensor arrays improves with more elements, yet increased number of elements leads to higher computational demands, cost and power consumption. Sparse arrays offer a cost-effective solution by utilizing only a subset of available elements. Each subset has a different effect on the performance properties of the array. This paper presents an unsupervised learning approach for sensor selection based on a deep generative modeling. The selection process is treated as a deterministic Markov Decision Process, where sensor subarrays arise as terminal states. The Generative Flow Network (GFlowNet) paradigm is employed to learn a distribution over actions based on the current state. Sampling from the aforementioned distribution ensures that the cumulative probability of reaching a terminal state is proportional to the sensing performance of the corresponding subset. The approach is applied for transmit beamforming where the performance of a subset is inversely proportional to the error between its corresponding beampattern and a desired beampattern. The method can generate multiple high-performing subsets by being trained on a small percentage of the possible subsets (less than 0.0001% of the possible subsets for the conducted experiments).