We built a framework for fragment-based 3D molecular design using reinforcement learning. Scales to more than 100 atoms.
Preprint: https://t.co/VgXt8Aozsu
w/ @alex_zhigalin, @A_Aspuru_Guzik
Watch the agent design a molecule by placing fragments from Imatinib:
ArXiv https://t.co/P0CRIZ3fbq: LSTMs with SMILES and SELFIES perform better at generating complex molecules than graph generative models. Confirming our own results.
Deep Learning can help solving many questions in Science. But how does it think about the world?
We analyze deep generative model's internal worldview for quantum designs: https://t.co/rxMgJXUG3D
Spearheaded by @FlamShepherd /w @ttttonywu@GuXuemei@albaclierta@A_Aspuru_Guzik
Stellarators are fusion reactors which use an optimized twisting magnetic cage to confine 100 million ℃ plasma. Here, I use AD (JAX) to optimize the coils which produce that cage.
Like ML in the 2010s, the stellarator community seems poised to adopt AD!
https://t.co/aC1sLKyIJo