The universe in a chip: high performance simulations of cosmic non-thermal accelerators

Ever since the advent of multi-wavelength astronomy, which extended observations beyond the narrow range accessible to the human eye, it has become clear that the Universe emits radiation across the entire electromagnetic spectrum. These observations revealed broadband spectral energy distributions that cannot be described by thermal equilibrium emission, such as a blackbody spectrum, but instead exhibit non-thermal features often characterized by power-law behaviour. This evidence indicates that non-thermal radiative processes, such as synchrotron emission, inverse Compton scattering, and high-energy hadronic interactions, play a fundamental role in astrophysical environments. In parallel, the discovery of cosmic rays demonstrated that the Universe hosts populations of highly energetic particles, raising fundamental questions about the mechanisms capable of accelerating them to relativistic energies. With the continuous advancement of observational techniques, it has become possible to resolve a wide range of astrophysical sources, including supernova remnants and relativistic jets, which are now recognized as key sites of particle acceleration and non-thermal emission. It is at this stage that numerical simulations started to play a key role in understanding system dynamics and particle acceleration pathways. One of the state-of-the-art tools for astrophysical simulations is PLUTO, a finite volume Godunov type code that evolves the relativistic magnetohydrodynamics equations on a discretized Eulerian gird. The code is equipped with a Lagrangian formalism that is able, by means of sub-grid prescriptions, to runtime relate the fluid conditions to the non-thermal spectrum of population of particles. The spectral evolution consistently includes the contributions of adiabatic expansion, synchrotron and inverse Compton emission, and is able to produce resolved synchrotron emission maps. In the present thesis I present a deeply revised implementation of such Lagrangian schemes, which, by means of modern C++, the Message Passing Interface library, and the use of the directivebased programming model OpenACC, enables Graphic Processing Units (GPUs) offloading to target the modern (pre-)exascale High Performance Computing infrastructures. GPUs, in fact, offer unprecedented speedup opportunity when we want to address the simulation of large scale astrophysical systems that pose boundless computational demand. The new implementation has been validated using reproducible test cases and has been deployed onto two EuroHPC JU computing facilities demonstrating excellent scaling capabilities up to 1024 parallel GPUs, as well a speedup of 6 times when solving that same benchmark test with 128 full GPU nodes (4 GPUs per node) against the same amount of full high-end CPU nodes (112 cores per node). The final part of the thesis presents a significant extension in range of physical applicability of the code, as I posed grate effort in completing the already present shock-driven particle acceleration prescription, by adding magnetic reconnection acceleration. Such implementation is able to evaluate the slope and the maximum energy of the non-thermal particle populations by sampling the values of plasma β and magnetization in the simulation, and it estimates the amount of magnetic energy feeding the acceleration, including the contribution given by the guide field. I applied the method to a 3D plasma column experiencing current-driven instability, where the current sheets developed fuel particle acceleration. The work provides the community with a faster, more efficient, and more capable code, able to advance our capability of testing particle acceleration models from relativistic magnetised shocks and magnetic reconnection in large scale astrophysical systems.