Sie verwenden einen veralteten Browser mit Sicherheitsschwachstellen und können die Funktionen dieser Webseite nicht nutzen.
Prof. Giuseppe Ciaccio, University of Genua
Freitag, 16.10.15, 10:15 Uhr
Haus 4, Raum 0.02
A superconducting magnet is an electromagnet made from windings of a multimaterial cable containing a superconducting wire at its core. When cooled to cryogenic temperature, the wire reaches the superconducting state and thus can conduct much larger electric currents than ordinary wire, creating intense magnetic fields without dissipating heat in the windings. Superconducting magnets are used in MRI machines in hospitals, and in scientific equipment such as NMR spectrometers, mass spectrometers and particle accelerators. In the simplest implementations, the windings take the shape of a classical solenoid. During regular service, a superconducting magnet may undergo a phenomenon called "quench". Due to an electric malfunction or mechanical stress or other reasons, a sudden transition from superconducting state to resistive state may occur in one or more regions of the wire. When this happens, the electromagnet starts dissipating its energy as heat in the resistive regions; if the cooling system is not properly dimensioned, such heat may trigger a transition to resistive state in other regions of the windings, so that more and more energy will be dissipated as heat, in a chain reaction that may ultimately lead to a catastrophic failure of the device.
A computer simulator of the whole magnet, able to reproduce a quench phenomenon in the space and time domains, would be very useful in the design stage, for properly dimensioning the various safety devices. The available simulators, however, use a very approximated model of the cable and thus may provide inaccurate results; moreover they have long execution time, do not run in parallel, and cannot be customized due to lack of source code.
For these reasons we have started developing a new simulator from scratch. We have modelled the device as a set of non-linear parabolic partial differential equations (PDEs) with piecewise continuous coefficients, where local linearization is performed in the discretization phase. For the solver, we have tried two well known finite difference discretization methods, namely, implicit Euler and Crank-Nicolson, solving the arising linear systems of equations with a simple parallel Jacobi method as a first attempt. From our preliminary results, implicit Euler emerges as substantially more efficient than the theoretically more precise Crank-Nicolson method, without incurring in a significantly higher discretization error. On an Infiniband cluster of Intel x86/64 dual-CPU nodes with 8 cores per CPU, the resolution shows a good speedup up to all the 40 cores made available to our tests, demonstrating the feasibility of the parallel approach.