I cannot honestly see an application for this. It's not clear what you are proposing, but I'll take a guess.
You have a description of a reservoir fluid and lots of associated data. This might include a breakdown of the fluid into its components, probably through gas chromatograph or similar, laboratory tests e.g., CVD, CCE for gases, assays on oil / condensate with details on paraffin, napthene and aromatic content for the C7+ fraction, perhaps some isotope analysis to help with source determination. You probably have samples, recombination or downhole (or both). You basically have a lot of data. Could you train a neural net to predict a general fluid category from that wealth of data? Absolutely. It would probably be trivial too as there are some pretty obvious input variables that you could apply to any standard data analysis approach. You don't really need an "A.I." approach.
Let me propose a different tact. The next step for many reservoir engineers once you've got this data is fluid characterisation. Let's build an equation of state that allows us to predict fluid properties at a wide range of temperature and pressure, from reservoir conditions through to surface conditions. Despite over a century of active research into EOS, across many industries (not just oil and gas) there still remains no single definitive solution. Tuning of EOS to match lab results is frequently required for reservoir simulation -- tuning mileage for facilities process engineers may vary. The late Mr. Baxendale used to say that the "pots and pans folk" don't know anything about tuning. Perhaps an "A.I." EOS could be trained to predict PVT properties for any arbitrary composition. This would be a huge breakthrough with value that goes well beyond the oil and gas industry.
Too hard...? Well, then the next thing you might notice is that a compositional simulator spends most of the computational time in the 'flash'. A flash calculation, when using a cubic EOS, uses a successive substitution process to iterate towards its solution. When you have a simulation model with hundreds of thousands or more cells (thanks Geomodellers) then that is a lot of flash calculations that are needed. So, if you can use "A.I." to speed up the flash calculation by perhaps providing a better initial estimate of vapour-liquid equilibrium, k-values or other similar parts of the flash, then that would be valuable as it would bring full compositional simulations closer to black-oil runtimes.
Then again, you might also observe that many compositional simulations depend on injection of a known secondary fluid that mixes with an in-situ reservoir fluid. You have two known fluids, and what varies between them is the fractional proportion of each. If you pre-compute the PVT tables for all possible fractions, then you could run a pseudo-compositional simulation just using look-up tables much like with a black oil model. The open source OPM Flow simulator has this capability through the PVTSOL keyword. I honestly don't know why this approach isn't used more frequently. Perhaps it is, and I'm just not aware of it. In other words, in this instance you might achieve something with "A.I." that has value. However, there are alternative workarounds already in existence that basic human intelligence could identify and utilise.
In summary, the value that "A.I." might bring when classifying reservoir fluids into five main fluids is negligible. However, if you can build a neural net to create a general equation of state that outperforms existing solutions, and can be calculated faster, then you could be onto a winner. Good luck.
------------------------------
Peter Kirkham
Project Commercialisation Manager
Twinza Oil
Australia
------------------------------