Knowledge Vault 5 /85 - CVPR 2023
Modeling Atoms to Address Our Climate Crisis
Larry Zitnick
< Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4o | Llama 3:

graph LR classDef chemistry fill:#f9d4d4, font-weight:bold, font-size:14px classDef catalysts fill:#d4f9d4, font-weight:bold, font-size:14px classDef ml fill:#d4d4f9, font-weight:bold, font-size:14px classDef challenges fill:#f9f9d4, font-weight:bold, font-size:14px classDef applications fill:#f9d4f9, font-weight:bold, font-size:14px classDef advice fill:#d4f9f9, font-weight:bold, font-size:14px A[Modeling Atoms to
Address Our Climate
Crisis] --> B[Chemistry: important,
deserves more attention 1] A --> C[1900s: Haber-Bosch process,
population growth 2] C --> D[Today: storing renewable energy,
catalysts 3] D --> E[Open Catalyst Project:
discover catalysts 4] E --> F[OC20, OC22:
140M training examples 5] E --> G[Graph networks estimate
atomic forces 6] G --> H[Equivariance: input-output
transform similarly 7] G --> I[Spherical harmonics model
atom orientations 8] A --> J[Challenges: accuracy,
uncertainty, more atoms 9] A --> K[Applications: batteries,
proteins, drugs, waste 10] A --> L[Explore new field:
identify problem 11] A --> M[Work across fields,
embrace science 12] class B,C,D chemistry class E,F catalysts class G,H,I ml class J challenges class K applications class L,M advice


1.- Chemistry is important and deserves more attention. Computer vision researchers are uniquely positioned to make progress in this field.

2.- In the early 1900s, chemists Haber and Bosch developed a process to create ammonia fertilizer, enabling population growth from 2 to 8 billion.

3.- Today, a key challenge is storing renewable energy. Chemical reactions can convert it to hydrogen or methane, but catalysts are needed.

4.- The Open Catalyst Project, a collaboration between Meta AI and CMU, aims to use ML to discover new catalysts.

5.- Datasets OC20 and OC22 contain 140M DFT training examples to help ML models generalize across elements and configurations.

6.- Graph neural networks are used to estimate atomic energy and forces. Orientation of atoms matters a lot.

7.- Equivariance is important - transform the input and the output should transform similarly. Steerable filters enable smoothly rotating filters.

8.- Spherical harmonics represent node embeddings to model orientations. Convolutions about the Y-axis can be done efficiently using Fourier transforms.

9.- The remaining challenges include improving accuracy, uncertainty estimation, modeling more atoms and time, and experimental validation of new materials.

10.- Other applications beyond renewable energy include direct air capture, batteries, protein folding, drug discovery, and waste cleanup.

11.- To explore a new field: make friends in that domain, identify an impactful problem together, then find a solution.

12.- Work across fields to maximize chances of new discoveries. Computer vision community should be open to new problems like AI for science.

Knowledge Vault built byDavid Vivancos 2024