Trial training first on predicting nominal parameters, transfer learning to fine-tune on real data

[lucpaoli]

Advantage:

• Potentially overcomes problems with nan / negative number inside logs
• Can train on known parameters, then introduce other molecules with unknown parameters

Perfect case-study! Train to predict SAFT VR Mie parameters for C1-C10, then see if it can continue to learn on dataset including isobutane, isopentane

[lucpaoli]

Update to this issue: Try training on published generated PCSAFT parameters from transformer paper, then swap out final MLP layer + SAFT head for SAFTVRMie & fine-tune on experimental / GERG data

[lucpaoli]

The PCSAFT parameters in the ETH paper probably aren't to be taken as gospel. Polar methane.

[lucpaoli]

This is a bad idea. Closing the issue.