Open lucpaoli opened 11 months ago
Also see 1800 compound PCPSAFT paper. Interesting discussion surrounding losses and weighting.
Especially interested about how to account for times when saturation pressure datapoints are above the predicted critical point for SAFT
2015 paper examines different data combinations for training, concluding that nothing more than sat. vap. pressure and sat. liq. densities are necessary. But they didn't look at degeneracy, and degeneracy has significant implications if the pure component parameters are used for mixtures (see pierre 2023 paper)
Which 2015 paper @MichaelGadaloff ?
I would like this to be part of a generic "architecture evaluation framework", that we can plug a general Flux.jl model into, train it, and see performance.
Could autodiff be used to determine sensitivity of model to hyperparameter selection? Not sure if that would be slower than a GLM or