matscholar-web v3: materials search should have a "predict" mode

materialsintelligence / matscholar-web

Code for the Materials Scholar website

http://matscholar.com

MIT License

9 stars 11 forks source link

Open computron opened 5 years ago

computron commented 5 years ago

In "predict" mode, if I searched for thermoelectrics, the app should:

rank materials by the dot product of word embedding/output embedding (same as Vahe's paper)
filter out materials where we have explicit mention of the thermoelectric application in a paper (i.e., any material in the current search results in the materials search app which shows up via entity search)

What will be left is just the novel predictions (assuming we are complete in abstract coverage)

jdagdelen commented 5 years ago

This may be somewhat computationally expensive (determining whether a term is in any abstract) and may not be something that can happen in real-time.

computron commented 5 years ago

I thought it's fast to search for text in abstracts? I did a text search for "thermoelectric" on matscholar just now and it was maybe 2 seconds?

jdagdelen commented 5 years ago

Yeah, you're right. We can probably do this real time.

ardunn commented 5 years ago

this seems to me like it should be a separate app in matscholar_web