ChatDFT

Inspiration

I read a fascinating preprint paper that was released a few days ago on chemRxiv link by Kevin Jablonka, a PhD student at EPFL. I talked with Kevin online on Thursday and thought it would be interesting to use TreeHacks as a time to pursue further research in this direction.

What it does

ChatDFT tells you if a high-entropy alloy is single or multiphase and then explains why it is that phase.

How we built it

I just used the openAI API for GPT-3, focusing on their text models.

Challenges we ran into

Trying to find adequate testing examples to construct my prompts with were difficult because I wanted my prompt to be balanced before it was used, but also it had to be inherently limited due to a lack of excessive characters needed for.

Accomplishments that we're proud of

What used to take 1257 data points to learn (Nature) now can be done in 9. Mostly. (I haven't done enough testing to be scientifically stated as SOTA)

What we learned

Integers are easier to work with for making categories for classification tasks than natural language.

What's next for ChatDFT

Expanding it to target other problems in chemistry, such as those with battery electrolytes or the activity of novel drugs.

Built With

• api
• openai