We made an AlphaGo like implementation for the card game Dominion. Certainly not the same number of cards but similar complexity. I have high confidence the same techniques would work for mtg. In fact possibly better as mtg doesn't lend to large search depths. Though possibly worse as there is more hidden information (though that depends on if the format has open deck lists and/or how much of the meta is provided to or trained by the nn)

I'm not aware of any good ML models for MTG. I'm just using off-the-shelf LLMs with a custom harness. It'd certainly be possible to do RLHF or something using the harness I've built, but it'd be expensive - anybody want to give me a few million dollars of OpenRouter credits so I can give it a shot?