Neutron stars are some of the most extreme objects in the Universe. Being very compact, with spin periods of a fraction of a second and magnetic fields that could reach up 10^15 G, they are capable of emitting radiation at all wavelengths from radio to X-rays to gamma-rays. Therefore, they are incredible laboratories to study fundamental physics and the properties of matter under extreme conditions. However, although about a billion neutron stars are expected to exist in the Milky Way, observational constraints limit us to only observing a few thousand. Pulsar population synthesis bridges this gap by simulating the full pulsar population and comparing it to the observed sample to constrain neutron-star physics. In this talk, we explore the possibility of using neural networks to estimate the parameters governing the properties of the entire pulsar population. For this purpose, we implement a population-synthesis framework able to simulate the dynamical and magneto-rotational evolution as well as the radio emission of isolated neutron stars and incorporate selection biases of typical radio surveys. By varying the parameters governing the natal kick-velocity distribution, the distribution of birth distances from the Galactic plane, the initial magnetic field and spin-period distributions, we create a dataset of mock observed radio pulsar populations that we use to train and validate a neural network for statistical inference. We describe how we use the resources at PIC and the HTCondor infrastructure to run simulations in parallel for this purpose. Our approach can help shed light on the physical mechanisms regulating the formation and evolution of neutron stars.