PHengLEI-NCCR/API/Physics/include/Flux_RoeEntropyCorrection.h

77 lines
4.2 KiB
C
Raw Normal View History

2024-10-14 09:32:17 +08:00
//+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
// PPPPP H H EEEEE N N GGGGG L EEEEE III +
// P P H H E NN N G L E I +
// PPPPP HHHHH EEEEE N N N G GG L EEEEE I +
// P H H E N N N G G L E I +
// P H H EEEEE N N GGGGG LLLLL EEEEE III +
//------------------------------------------------------------------------+
// Platform for Hybrid Engineering Simulation of Flows +
// China Aerodynamics Research and Development Center +
// (C) Copyright, Since 2010 +
//+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
//! @file Flux_RoeEntropyCorrection.h
//! @brief This file defines the entropy correction (fix) method in Roe inviscid flux method.
//! @author Bell, Liu Jian.
#pragma once
#include "LIB_Macro.h"
#include "Math_BasisFunction.h"
#ifdef USE_GMRESSOLVER
#include "Sacado.hpp"
#endif
namespace PHSPACE
{
#ifdef USE_GMRESSOLVER
//! GMRESAD
typedef Sacado::Fad::DFad<RDouble> ADReal;
//! The entropy correction (fix) in Roe scheme of inviscid flux, which is used to limit the magnitude of the three eigenvalue.\n
//! @param[in] eigv1,2,3 The three eigenvalue, which are V/V+c/V-c on the face. They will be corrected and returned.\n
//! @param[in] vn The normal velocity on the face.\n
//! @param[in] absVel The absolute velocity on the face.\n
//! @param[in] RoeEntropyCorrectionMethod The Roe entropy correction method.\n
//! @param[in] entropyCorrectionCoef The Roe entropy correction coefficient.\n
template <typename T>
LIB_EXPORT void Flux_RoeEntropyCorrection(T &eigv1, T &eigv2, T &eigv3, T pressureCoeffL, T pressureCoeffR,
T absVel, T vn, T cm, int RoeEntropyCorrectionMethod, T acousticEigenLimit, T convectEigenLimit);
//! Harten type of entropy correction (fix).\n
//! @param[in] eigenValue The original eigenvalue, which may be invalid.
//! @param[in] eigenValueLimit The epsilon in the Harten formula.
//! @return The new eigenvalue which has been limited.
template <typename T>
inline T Harten(T eigenValue, T eigenValueLimit)
{
T eigenValueNew = abs(eigenValue);
if(eigenValueNew < eigenValueLimit)
{
eigenValueNew = 0.5 * (eigenValue * eigenValue + eigenValueLimit * eigenValueLimit) / eigenValueLimit;
}
return eigenValueNew;
}
#else
//! The entropy correction (fix) in Roe scheme of inviscid flux, which is used to limit the magnitude of the three eigenvalue.\n
//! @param[in] eigv1,2,3 The three eigenvalue, which are V/V+c/V-c on the face. They will be corrected and returned.\n
//! @param[in] vn The normal velocity on the face.\n
//! @param[in] absVel The absolute velocity on the face.\n
//! @param[in] RoeEntropyCorrectionMethod The Roe entropy correction method.\n
//! @param[in] entropyCorrectionCoef The Roe entropy correction coefficient.\n
LIB_EXPORT void Flux_RoeEntropyCorrection(RDouble &eigv1, RDouble &eigv2, RDouble &eigv3, RDouble pressureCoeffL, RDouble pressureCoeffR,
RDouble absVel, RDouble vn, RDouble cm, int RoeEntropyCorrectionMethod, RDouble acousticEigenLimit, RDouble convectEigenLimit);
//! Harten type of entropy correction (fix).\n
//! @param[in] eigenValue The original eigenvalue, which may be invalid.
//! @param[in] eigenValueLimit The epsilon in the Harten formula.
//! @return The new eigenvalue which has been limited.
inline RDouble Harten(RDouble eigenValue, RDouble eigenValueLimit)
{
RDouble eigenValueNew = ABS(eigenValue);
if(eigenValueNew < eigenValueLimit)
{
eigenValueNew = 0.5 * (eigenValue * eigenValue + eigenValueLimit * eigenValueLimit) / eigenValueLimit;
}
return eigenValueNew;
}
#endif
}