Open research towards the discovery of room-temperature superconductors.
Discover ways to transform this asset
POST /speech/from-post
In a simplified version of the full end-to-end pipeline, we attempt to use MatterGen, recently released by Microsoft, to generate novel materials and test them for superconductivity.
MatterGen allows you to explore chemical families with a simple input like:
Bi-Ca-Cu-Fe-O
Bi-Ca-Cu-Sr-O
Ba-Pb-Sb-O
Ba-Cu-La
Ca-Cu
Bi-K-O
Many of these families are known to have superconductors already, so exploring new configurations within them seems like a good place to start.
To generate a batch of materials, 32 in our case, it takes about 45 minutes. Evaluating them for superconductivity takes around the same amount of time (NVIDIA T4).
The vast majority of materials generated are classified as non-superconductors. This is good! It means we may have a model that is decently calibrated to the rarity of superconductors. We are also finding plenty of low temp (sub 10 K) superconductors. It's good to see that even within a single chemistry, we can predict both superconducting and non-superconducting.
We may have discovered a new, high-Tc superconductor! gen_8 from Ba-Cu-F-Sr-O is estimated to have a Tc near 113 K. That said, note how uncertain the model is. We've seen this before during evaluation, but it's generally a good sign for true superconductors, it just means it's outside it's training distribution.
Visualizing predictions for a novel material in the Ba-Cu-F-Sr-O family, estimated to have at Tc of 113 K.
You can find the CIF file for the crystal here:
Unconfirmed new high-Tc superconductor generated by MatterGen. Estimated to have a critical temperature around 113 K.
More analysis is required to confirm its uniqueness. With a Tc of 113 K (-160.2°C), this material could reach it's superconducting state with commonly available liquid nitrogen at -196°C.
Bi-Ca-Cu-Sr-O gen_19 with a Tc estimate of 38.7 K
Bi-Ca-Cu-Sr-O gen_3 with Tc estimate of 104.0K
As far as I know, MatterGen only works to generate materials with up to 5 elements. This may be okay as this is still such a large search space, but I've seen superconductors composed of more than 5. There are also many with fewer than 5 unique elements too.
MatterGen has functionality where you can fine-tune the model to draw out generations that optimize for some property. We have all the tools needed to take 3DSC and train a version of MatterGen that we can use to directly generate high-Tc materials. Having the classifier model will help validate our generations.
Further work on stability of generated materials is also needed.
Discover other posts like this one
So far a really interesting paper. Published in 2018. Adding some informal notes and interesting findings here. Finding out how much literature is based on this study.
Using what we learned when trying to use the MLFF's latent space for Tc prediction, there's a way we can simplify things for the prediction model and give it a better change of picking up on the signa
Careful evaluation of the classifier model is important so that we can truly understand the capabilities and performance of a Tc predicting model. Particularly important to us is the ability for the m
Building on 's work with the fine-tuned MatterGen model, we evaluated 400 candidate materials for superconductivity using the Tc classification model. Prior work on fine-tuning: