Emily Boudreaux
cd191cff23
Switching to an Engine + solver design. Also brought xxHash and Eigen in. Working on QSE and Culling.
384 lines
16 KiB
C++
384 lines
16 KiB
C++
#include "gridfire/solver/solver.h"
|
||
#include "gridfire/engine/engine_graph.h"
|
||
#include "gridfire/network.h"
|
||
|
||
#include "fourdst/composition/atomicSpecies.h"
|
||
#include "fourdst/composition/composition.h"
|
||
#include "fourdst/config/config.h"
|
||
|
||
#include "Eigen/Dense"
|
||
#include "unsupported/Eigen/NonLinearOptimization"
|
||
|
||
#include <boost/numeric/odeint.hpp>
|
||
|
||
#include <vector>
|
||
#include <unordered_map>
|
||
#include <string>
|
||
#include <stdexcept>
|
||
|
||
#include "quill/LogMacros.h"
|
||
|
||
namespace gridfire::solver {
|
||
|
||
NetOut QSENetworkSolver::evaluate(const NetIn &netIn) {
|
||
using state_type = boost::numeric::ublas::vector<double>;
|
||
using namespace boost::numeric::odeint;
|
||
|
||
NetOut postIgnition = initializeNetworkWithShortIgnition(netIn);
|
||
|
||
constexpr double abundance_floor = 1.0e-30;
|
||
std::vector<double>Y_sanitized_initial;
|
||
Y_sanitized_initial.reserve(m_engine.getNetworkSpecies().size());
|
||
|
||
LOG_DEBUG(m_logger, "Sanitizing initial abundances with a floor of {:0.3E}...", abundance_floor);
|
||
for (const auto& species : m_engine.getNetworkSpecies()) {
|
||
double molar_abundance = 0.0;
|
||
if (postIgnition.composition.contains(species)) {
|
||
molar_abundance = postIgnition.composition.getMolarAbundance(std::string(species.name()));
|
||
}
|
||
|
||
if (molar_abundance < abundance_floor) {
|
||
molar_abundance = abundance_floor;
|
||
}
|
||
Y_sanitized_initial.push_back(molar_abundance);
|
||
}
|
||
const double T9 = netIn.temperature / 1e9; // Convert temperature from Kelvin to T9 (T9 = T / 1e9)
|
||
const double rho = netIn.density; // Density in g/cm^3
|
||
|
||
const auto indices = packSpeciesTypeIndexVectors(Y_sanitized_initial, T9, rho);
|
||
Eigen::VectorXd Y_QSE;
|
||
try {
|
||
Y_QSE = calculateSteadyStateAbundances(Y_sanitized_initial, T9, rho, indices);
|
||
} catch (const std::runtime_error& e) {
|
||
LOG_ERROR(m_logger, "Failed to calculate steady state abundances. Aborting QSE evaluation.");
|
||
m_logger->flush_log();
|
||
throw std::runtime_error("Failed to calculate steady state abundances: " + std::string(e.what()));
|
||
}
|
||
|
||
state_type YDynamic_ublas(indices.dynamicSpeciesIndices.size() + 1);
|
||
for (size_t i = 0; i < indices.dynamicSpeciesIndices.size(); ++i) {
|
||
YDynamic_ublas(i) = Y_sanitized_initial[indices.dynamicSpeciesIndices[i]];
|
||
}
|
||
YDynamic_ublas(indices.dynamicSpeciesIndices.size()) = 0.0; // Placeholder for specific energy rate
|
||
|
||
const RHSFunctor rhs_functor(m_engine, indices.dynamicSpeciesIndices, indices.QSESpeciesIndices, Y_QSE, T9, rho);
|
||
const auto stepper = make_controlled<runge_kutta_dopri5<state_type>>(1.0e-8, 1.0e-8);
|
||
|
||
size_t stepCount = integrate_adaptive(
|
||
stepper,
|
||
rhs_functor,
|
||
YDynamic_ublas,
|
||
0.0, // Start time
|
||
netIn.tMax,
|
||
netIn.dt0
|
||
);
|
||
|
||
std::vector<double> YFinal = Y_sanitized_initial;
|
||
for (size_t i = 0; i <indices.dynamicSpeciesIndices.size(); ++i) {
|
||
YFinal[indices.dynamicSpeciesIndices[i]] = YDynamic_ublas(i);
|
||
}
|
||
for (size_t i = 0; i < indices.QSESpeciesIndices.size(); ++i) {
|
||
YFinal[indices.QSESpeciesIndices[i]] = Y_QSE(i);
|
||
}
|
||
|
||
const double finalSpecificEnergyRate = YDynamic_ublas(indices.dynamicSpeciesIndices.size());
|
||
|
||
// --- Marshal output variables ---
|
||
std::vector<std::string> speciesNames(m_engine.getNetworkSpecies().size());
|
||
std::vector<double> finalMassFractions(m_engine.getNetworkSpecies().size());
|
||
double massFractionSum = 0.0;
|
||
|
||
for (size_t i = 0; i < speciesNames.size(); ++i) {
|
||
const auto& species = m_engine.getNetworkSpecies()[i];
|
||
speciesNames[i] = species.name();
|
||
finalMassFractions[i] = YFinal[i] * species.mass(); // Convert from molar abundance to mass fraction
|
||
massFractionSum += finalMassFractions[i];
|
||
}
|
||
for (auto& mf : finalMassFractions) {
|
||
mf /= massFractionSum; // Normalize to get mass fractions
|
||
}
|
||
|
||
fourdst::composition::Composition outputComposition(speciesNames, finalMassFractions);
|
||
NetOut netOut;
|
||
netOut.composition = outputComposition;
|
||
netOut.energy = finalSpecificEnergyRate; // Specific energy rate
|
||
netOut.num_steps = stepCount;
|
||
return netOut;
|
||
}
|
||
|
||
dynamicQSESpeciesIndices QSENetworkSolver::packSpeciesTypeIndexVectors(
|
||
const std::vector<double>& Y,
|
||
const double T9,
|
||
const double rho
|
||
) const {
|
||
constexpr double timescaleCutoff = 1.0e-5;
|
||
constexpr double abundanceCutoff = 1.0e-15;
|
||
|
||
LOG_INFO(m_logger, "Partitioning species using T9={:0.2f} and ρ={:0.2e}", T9, rho);
|
||
LOG_INFO(m_logger, "Timescale Cutoff: {:.1e} s, Abundance Cutoff: {:.1e}", timescaleCutoff, abundanceCutoff);
|
||
|
||
std::vector<size_t>dynamicSpeciesIndices; // Slow species that are not in QSE
|
||
std::vector<size_t>QSESpeciesIndices; // Fast species that are in QSE
|
||
|
||
std::unordered_map<fourdst::atomic::Species, double> speciesTimescale = m_engine.getSpeciesTimescales(Y, T9, rho);
|
||
|
||
for (size_t i = 0; i < m_engine.getNetworkSpecies().size(); ++i) {
|
||
const auto& species = m_engine.getNetworkSpecies()[i];
|
||
const double timescale = speciesTimescale[species];
|
||
const double abundance = Y[i];
|
||
|
||
if (std::isinf(timescale) || abundance < abundanceCutoff || timescale <= timescaleCutoff) {
|
||
QSESpeciesIndices.push_back(i);
|
||
} else {
|
||
dynamicSpeciesIndices.push_back(i);
|
||
}
|
||
}
|
||
LOG_INFO(m_logger, "Partitioning complete. Dynamical species: {}, QSE species: {}.", dynamicSpeciesIndices.size(), QSESpeciesIndices.size());
|
||
LOG_INFO(m_logger, "Dynamic species: {}", [dynamicSpeciesIndices](const DynamicEngine& engine_wrapper) -> std::string {
|
||
std::string result;
|
||
int count = 0;
|
||
for (const auto& i : dynamicSpeciesIndices) {
|
||
result += std::string(engine_wrapper.getNetworkSpecies()[i].name());
|
||
if (count < dynamicSpeciesIndices.size() - 2) {
|
||
result += ", ";
|
||
} else if (count == dynamicSpeciesIndices.size() - 2) {
|
||
result += " and ";
|
||
}
|
||
count++;
|
||
}
|
||
return result;
|
||
}(m_engine));
|
||
LOG_INFO(m_logger, "QSE species: {}", [QSESpeciesIndices](const DynamicEngine& engine_wrapper) -> std::string {
|
||
std::string result;
|
||
int count = 0;
|
||
for (const auto& i : QSESpeciesIndices) {
|
||
result += std::string(engine_wrapper.getNetworkSpecies()[i].name());
|
||
if (count < QSESpeciesIndices.size() - 2) {
|
||
result += ", ";
|
||
} else if (count == QSESpeciesIndices.size() - 2) {
|
||
result += " and ";
|
||
}
|
||
count++;
|
||
}
|
||
return result;
|
||
}(m_engine));
|
||
return {dynamicSpeciesIndices, QSESpeciesIndices};
|
||
}
|
||
|
||
Eigen::VectorXd QSENetworkSolver::calculateSteadyStateAbundances(
|
||
const std::vector<double> &Y,
|
||
const double T9,
|
||
const double rho,
|
||
const dynamicQSESpeciesIndices &indices
|
||
) const {
|
||
std::vector<double> Y_dyn;
|
||
Eigen::VectorXd Y_qse_initial(indices.QSESpeciesIndices.size());
|
||
for (const auto& i : indices.dynamicSpeciesIndices) { Y_dyn.push_back(Y[i]); }
|
||
for (size_t i = 0; i < indices.QSESpeciesIndices.size(); ++i) {
|
||
Y_qse_initial(i) = Y[indices.QSESpeciesIndices[i]];
|
||
if (Y_qse_initial(i) < 1.0e-99) { Y_qse_initial(i) = 1.0e-99; }
|
||
}
|
||
|
||
Eigen::VectorXd v_qse = Y_qse_initial.array().log();
|
||
|
||
EigenFunctor<double> qse_problem(m_engine, Y_dyn, indices.dynamicSpeciesIndices, indices.QSESpeciesIndices, T9, rho);
|
||
LOG_INFO(m_logger, "--- QSE Pre-Solve Diagnostics ---");
|
||
|
||
Eigen::VectorXd f_initial(indices.QSESpeciesIndices.size());
|
||
qse_problem(v_qse, f_initial);
|
||
LOG_INFO(m_logger, "Initial Guess ||f||: {:0.4e}", f_initial.norm());
|
||
|
||
Eigen::MatrixXd J_initial(indices.QSESpeciesIndices.size(), indices.QSESpeciesIndices.size());
|
||
qse_problem.df(v_qse, J_initial);
|
||
const Eigen::JacobiSVD<Eigen::MatrixXd> svd(J_initial);
|
||
double cond = svd.singularValues().maxCoeff() / svd.singularValues().minCoeff();
|
||
LOG_INFO(m_logger, "Initial Jacobian Condition Number: {:0.4e}", cond);
|
||
LOG_INFO(m_logger, "Starting QSE solve...");
|
||
|
||
Eigen::HybridNonLinearSolver<EigenFunctor<double>> solver(qse_problem);
|
||
solver.parameters.xtol = 1.0e-8; // Set tolerance
|
||
|
||
// 5. Run the solver. It will modify v_qse in place.
|
||
const int eigenStatus = solver.solve(v_qse);
|
||
|
||
// 6. Check for convergence and return the result
|
||
if(eigenStatus != Eigen::Success) {
|
||
LOG_WARNING(m_logger, "--- QSE SOLVER FAILED ---");
|
||
LOG_WARNING(m_logger, "Eigen status code: {}", eigenStatus);
|
||
LOG_WARNING(m_logger, "Iterations performed: {}", solver.iter);
|
||
|
||
// Log the final state that caused the failure
|
||
Eigen::VectorXd Y_qse_final_fail = v_qse.array().exp();
|
||
for(long i=0; i<v_qse.size(); ++i) {
|
||
LOG_WARNING(m_logger, "Final v_qse[{}]: {:0.4e} -> Y_qse[{}]: {:0.4e}", i, v_qse(i), i, Y_qse_final_fail(i));
|
||
}
|
||
|
||
// Log the residual at the final state
|
||
Eigen::VectorXd f_final(indices.QSESpeciesIndices.size());
|
||
qse_problem(v_qse, f_final);
|
||
LOG_WARNING(m_logger, "Final ||f||: {:0.4e}", f_final.norm());
|
||
|
||
throw std::runtime_error("Eigen QSE solver did not converge.");
|
||
}
|
||
LOG_INFO(m_logger, "Eigen QSE solver converged in {} iterations.", solver.iter);
|
||
|
||
return v_qse.array().exp();
|
||
|
||
}
|
||
|
||
NetOut QSENetworkSolver::initializeNetworkWithShortIgnition(const NetIn &netIn) const {
|
||
const auto ignitionTemperature = m_config.get<double>(
|
||
"gridfire:solver:QSE:ignition:temperature",
|
||
2e8
|
||
); // 0.2 GK
|
||
const auto ignitionDensity = m_config.get<double>(
|
||
"gridfire:solver:QSE:ignition:density",
|
||
1e6
|
||
); // 1e6 g/cm^3
|
||
const auto ignitionTime = m_config.get<double>(
|
||
"gridfire:solver:QSE:ignition:tMax",
|
||
1e-7
|
||
); // 0.1 μs
|
||
const auto ignitionStepSize = m_config.get<double>(
|
||
"gridfire:solver:QSE:ignition:dt0",
|
||
1e-15
|
||
); // 1e-15 seconds
|
||
|
||
LOG_INFO(
|
||
m_logger,
|
||
"Igniting network with T={:<5.3E}, ρ={:<5.3E}, tMax={:<5.3E}, dt0={:<5.3E}...",
|
||
ignitionTemperature,
|
||
ignitionDensity,
|
||
ignitionTime,
|
||
ignitionStepSize
|
||
);
|
||
|
||
NetIn preIgnition = netIn;
|
||
preIgnition.temperature = ignitionTemperature;
|
||
preIgnition.density = ignitionDensity;
|
||
preIgnition.tMax = ignitionTime;
|
||
preIgnition.dt0 = ignitionStepSize;
|
||
|
||
DirectNetworkSolver ignitionSolver(m_engine);
|
||
NetOut postIgnition = ignitionSolver.evaluate(preIgnition);
|
||
LOG_INFO(m_logger, "Network ignition completed in {} steps.", postIgnition.num_steps);
|
||
return postIgnition;
|
||
}
|
||
|
||
void QSENetworkSolver::RHSFunctor::operator()(
|
||
const boost::numeric::ublas::vector<double> &YDynamic,
|
||
boost::numeric::ublas::vector<double> &dYdtDynamic,
|
||
double t
|
||
) const {
|
||
// --- Populate the slow / dynamic species vector ---
|
||
std::vector<double> YFull(m_engine.getNetworkSpecies().size());
|
||
for (size_t i = 0; i < m_dynamicSpeciesIndices.size(); ++i) {
|
||
YFull[m_dynamicSpeciesIndices[i]] = YDynamic(i);
|
||
}
|
||
|
||
// --- Populate the QSE species vector ---
|
||
for (size_t i = 0; i < m_QSESpeciesIndices.size(); ++i) {
|
||
YFull[m_QSESpeciesIndices[i]] = m_Y_QSE(i);
|
||
}
|
||
|
||
auto [full_dYdt, specificEnergyRate] = m_engine.calculateRHSAndEnergy(YFull, m_T9, m_rho);
|
||
|
||
dYdtDynamic.resize(m_dynamicSpeciesIndices.size() + 1);
|
||
for (size_t i = 0; i < m_dynamicSpeciesIndices.size(); ++i) {
|
||
dYdtDynamic(i) = full_dYdt[m_dynamicSpeciesIndices[i]];
|
||
}
|
||
dYdtDynamic[m_dynamicSpeciesIndices.size()] = specificEnergyRate;
|
||
}
|
||
|
||
NetOut DirectNetworkSolver::evaluate(const NetIn &netIn) {
|
||
namespace ublas = boost::numeric::ublas;
|
||
namespace odeint = boost::numeric::odeint;
|
||
using fourdst::composition::Composition;
|
||
|
||
const double T9 = netIn.temperature / 1e9; // Convert temperature from Kelvin to T9 (T9 = T / 1e9)
|
||
const unsigned long numSpecies = m_engine.getNetworkSpecies().size();
|
||
|
||
const auto absTol = m_config.get<double>("gridfire:solver:DirectNetworkSolver:absTol", 1.0e-8);
|
||
const auto relTol = m_config.get<double>("gridfire:solver:DirectNetworkSolver:relTol", 1.0e-8);
|
||
|
||
size_t stepCount = 0;
|
||
|
||
RHSFunctor rhsFunctor(m_engine, T9, netIn.density);
|
||
JacobianFunctor jacobianFunctor(m_engine, T9, netIn.density);
|
||
|
||
ublas::vector<double> Y(numSpecies + 1);
|
||
|
||
for (size_t i = 0; i < numSpecies; ++i) {
|
||
const auto& species = m_engine.getNetworkSpecies()[i];
|
||
try {
|
||
Y(i) = netIn.composition.getMolarAbundance(std::string(species.name()));
|
||
} catch (const std::runtime_error) {
|
||
LOG_DEBUG(m_logger, "Species '{}' not found in composition. Setting abundance to 0.0.", species.name());
|
||
Y(i) = 0.0;
|
||
}
|
||
}
|
||
Y(numSpecies) = 0.0;
|
||
|
||
const auto stepper = odeint::make_controlled<odeint::rosenbrock4<double>>(absTol, relTol);
|
||
stepCount = odeint::integrate_adaptive(
|
||
stepper,
|
||
std::make_pair(rhsFunctor, jacobianFunctor),
|
||
Y,
|
||
0.0,
|
||
netIn.tMax,
|
||
netIn.dt0
|
||
);
|
||
|
||
std::vector<double> finalMassFractions(numSpecies);
|
||
for (size_t i = 0; i < numSpecies; ++i) {
|
||
const double molarMass = m_engine.getNetworkSpecies()[i].mass();
|
||
finalMassFractions[i] = Y(i) * molarMass; // Convert from molar abundance to mass fraction
|
||
if (finalMassFractions[i] < MIN_ABUNDANCE_THRESHOLD) {
|
||
finalMassFractions[i] = 0.0;
|
||
}
|
||
}
|
||
|
||
std::vector<std::string> speciesNames;
|
||
speciesNames.reserve(numSpecies);
|
||
for (const auto& species : m_engine.getNetworkSpecies()) {
|
||
speciesNames.push_back(std::string(species.name()));
|
||
}
|
||
|
||
Composition outputComposition(speciesNames, finalMassFractions);
|
||
outputComposition.finalize(true);
|
||
|
||
NetOut netOut;
|
||
netOut.composition = std::move(outputComposition);
|
||
netOut.energy = Y(numSpecies); // Specific energy rate
|
||
netOut.num_steps = stepCount;
|
||
|
||
return netOut;
|
||
}
|
||
|
||
void DirectNetworkSolver::RHSFunctor::operator()(
|
||
const boost::numeric::ublas::vector<double> &Y,
|
||
boost::numeric::ublas::vector<double> &dYdt,
|
||
double t
|
||
) const {
|
||
const std::vector<double> y(Y.begin(), m_numSpecies + Y.begin());
|
||
auto [dydt, eps] = m_engine.calculateRHSAndEnergy(y, m_T9, m_rho);
|
||
dYdt.resize(m_numSpecies + 1);
|
||
std::ranges::copy(dydt, dydt.begin());
|
||
dYdt(m_numSpecies) = eps;
|
||
}
|
||
|
||
void DirectNetworkSolver::JacobianFunctor::operator()(
|
||
const boost::numeric::ublas::vector<double> &Y,
|
||
boost::numeric::ublas::matrix<double> &J,
|
||
double t,
|
||
boost::numeric::ublas::vector<double> &dfdt
|
||
) const {
|
||
J.resize(m_numSpecies + 1, m_numSpecies + 1);
|
||
J.clear();
|
||
for (int i = 0; i < m_numSpecies + 1; ++i) {
|
||
for (int j = 0; j < m_numSpecies + 1; ++j) {
|
||
J(i, j) = m_engine.getJacobianMatrixEntry(i, j);
|
||
}
|
||
}
|
||
}
|
||
} |