Repository navigation
Fix Python compatible StateHD<D>s and respective dual numbers #92
Description
For the python equation of state, users can define their own helmholtz_energy function that takes a StateHD object as input. In the function, properties of the state can be accessed, e.g. one can write state.moles.
To make this work, there has to be a non-generic state object for each (hyper) dual number.
Currently, missing dual numbers are simply defined in feos-core. Using these definitions, for each a non-generic StateHD version is built.
There are currently two issues:
- simply defining the dual numbers is not enough - we need arithmetic operations, etc. from
num-dual(which currently does not export all variants we need to Python) - we do not call
impl_state_hd!(which implement the getters) for all states that are built. Hence, there will be errors in Python when one tries to call the getters.
To fix this, we
- have to properly export all dual numbers that are needed from
num-dual. (Check class names in macros!) - properly build all state definitions
- build state definition and implementation for
HelmholtzEnergyDualin single macro?
Activity
prehner commentedon Dec 20, 2022
Since we have to deal with this problem now anyways: Is there any way we can make the implementation of the Python FFI for the dual numbers more ergonomic? It's virtually impossible to make a
Box<dyn DualNum<f64>>using aPyDualNumber<PyObject>and doing a (manually implemented) type check for every operation we have also seems unfeasible.Aside from that the steps that you outline seem doable. Maybe expose ALL dual numbers that are created in the num-dual Python FFI and then in feos we can just build the States that we need. E.g., currently num-dual builds dual numbers with up to 10 derivatives but we would only need them for up to 2 in feos.
(For more than 10 derivatives backward AD should be seriously considered anyways)
g-bauer commentedon Dec 20, 2022
These are things to consider:
num-dualmodule (the Python package)For
feoswe just have to make sure that all variants we need inHelmholtzEnergyare build and public in thepython(Rust) module. But I agree, we should investigate if any improvements to the way we build the bindings can be made. This is, however, an issue fornum-dual.In
feos, currently two macros are used. One implements the getters for the states that are at the moment defined manually, the second one implements theHelmholtzEnergytrait. It should be possible to do that all automatically given a list of dual numbers.