allanleal
commented
3 years ago Change your alphar method as follows:
template<typename TType, typename RhoType>
auto alphar(const TType &T, const RhoType& rho) const -> TType {
auto rhotot = std::accumulate(std::begin(rho), std::end(rho), static_cast<typename RhoType::value_type>(0.0));
auto Psiminus = -log(1.0 - b * rhotot);
auto Psiplus = rhotot;
return Psiminus - a / (R * T) * Psiplus;
}Changes
-
Make sure you specify the return type as
TType. Reason: Operations withautodiff::dualproduce a compile-time expression. This expression may contain references to objects in the function scope. By specifying the return type toTType(orautodiff::dual), you force the expression to be evaluated at the exit of the function. This is very much like what happens to Eigen as well (try to return an Eigen expression involving vectors/matrices constructed inside a function!). There is no way around this -- well, there is: using shared pointers, but the performance would be ridiculous for forward differentiation. -
I had to use
static_cast<typename RhoType::value_type>(0.0)instead of your(RhoType::value_type)0.0.
This compiled fine with clang++ v11 and g++ v7.3.
I would be interested to learn more about your equation of state based on residual Helmholtz energy! ;) Can this be applied to any solution (e.g., electrolyte solutions)?
ianhbell
I have written a C++ library that allows for evaluation of thermodynamic equation of state written in terms of residual Helmholtz energy (only, not derivatives): https://github.com/ianhbell/teqp . Derivatives are taken with complex derivatives for single derivatives, and multicomplex derivatives with higher (and cross) derivatives. This works great with multicomplex differentiation, but it is not very fast, so I wanted to see if automatic differentiation would be faster. The core function needs to accept arbitrary numerical types as template arguments.
The below is a simple test case showing how I would like to use
autodiff, in analogy to what I did with the complex differentiation tools. The generic model implements the functionalpharand here I am trying to take the derivative of the first argument, which is scalar, as a first test.Visual studio gives voluminous compilation errors (below). I suspect it has something to do with the fact that the function I am wrapping in the lambda is an instance method, though I admit I haven't the foggiest idea how to workaround this issue because the lambda function approach is usually a success in other cases.