177,344 research outputs found

    Thermodynamics modeling of deep learning systems for a temperature based filter pruning technique

    No full text
    We analyse the dynamics of convolutional filters' parameters of a convolutional neural networks during and after training, via a thermodynamic analogy which allows for a sound definition of temperature. We show that removing high temperature filters has a minor effect on the performance of the model, while removing low temperature filters influences majorly both accuracy and loss decay. This result could be exploited to implement a temperature-based pruning technique for the filters and to determine efficiently the crucial filters for an effective learning

    Subgrid modeling of intrinsic instabilities in premixed flame propagation

    No full text
    This work is devoted to the investigation and subgrid-scale modeling of intrinsic flame instabilities occurring in the propagation of a deflagration wave. Such instabilities, of hydrodynamic and thermodiffusive origin, are expected to be of particular relevance in recent technological trends such as in the use of hydrogen as a clean energy carrier or as a secondary fuel in hydrogen enriched combustion. A dedicated set of direct numerical simulations is presented and used, in conjunction with coherent literature results, in order to develop scaling arguments for the propagation speed of self-wrinkled flames which are also supported by the outcomes of a weakly non-linear model, namely the Sivashinsky equation. The observed scaling is based on the definition of the number of unstable wavelengths in a reference hydrodynamic lengthscale, in other words the ratio between the neutral or cutoff lengthscale of intrinsic instabilities and the lateral domain of a planar flame. The scalings are then employed to develop an algebraic model for the wrinkling factor in the context of a flame surface density closure approach. An a-priori analysis shows that the model correctly captures the flame wrinkling caused by intrinsic instability at sub grid level. A strategy to include the developed self-wrinkling model in the context of a turbulent combustion model is finally discussed on the basis of the turbulence induced cut-off concept
    corecore