name: materialdiscovery ## Accelerating material discovery with machine learning
.center[![:scale 80%](/assets/images/slides/materials/materials_sample.jpg)] --- ## Accelerating material discovery with machine learning ### Motivation .center[
.cite[(Lawrence Zitnick et al., 2020)]
] * Renewable energy can be used to transform water into hydrogen or methane and back to electricity * However, current electrocatalysts are not sufficiently energy-efficient (35 % for round-trip AC to AC) .references[ * Lawrenece Zitnick et al. [An introduction to electrocatalyst design using machine learning for renewable energy storage](https://arxiv.org/abs/2010.09435). arXiv:2010.09435, 2020. ] --- ## Why machine learning? ### Traditional electrocatalyst design .context[Current electrocatalyst are only up to 35 % energy efficient] .right-column-66[.center[![:scale 90%](/assets/images/slides/materials/activelearning_noml.png)]] -- .left-column-33[ A _relaxation_ of propane (C3H8) on a copper (Cu) surface.
.center[![:scale 100%](/assets/images/slides/materials/relaxation_crop.gif)] ] --- count: false ## Why machine learning? ### Traditional electrocatalyst design .context[Current electrocatalyst are only up to 35 % energy efficient] .right-column-66[.center[![:scale 90%](/assets/images/slides/materials/activelearning_noml.png)]] .left-column-33[ * Density Functional Theory is used to estimate the energy of a catalyst-molecule structure * DFT scales with $O(n^3)$ with the number of electrons * The calculations for one structure take hours or days * There are combinatorially many possible candidate materials ] --- ## Why machine learning? ### ML world model .context[Physical models are computationally too expensive for fast discovery.] .right-column-66[.center[![:scale 90%](/assets/images/slides/materials/activelearning_ml.png)]] -- .left-column-33[ * Data from physical models can be used to train ML-based approximators * ML models can be used to more rapidly evaluate candidate materials ] -- Can we do better? --- ## Why machine learning? ### RL-based exploratory policy .context[Using ML to only score candidate materials provides only _linear_ gains.] .right-column-66[.center[![:scale 90%](/assets/images/slides/materials/activelearning_agent.png)]] -- .left-column-33[ * We can train ML models to more efficiently search the space of candidate materials * An RL agent could exploit the structure of the search space ] Promising results: [GFlowNet](https://arxiv.org/abs/2106.04399) (Bengio et al., 2021) --- ## Accelerating scientific discovery ### Summary .right-column[.center[![:scale 90%](/assets/images/slides/materials/activelearning_agent.png)]] .left-column[ * World model: graph neural networks (GNN), capable of incorporating invariances and equivariances that preserve physical properties * Exploratory agent: RL-based algorithms capable of learning the structure of the _world_ to propose diverse, high-reward candidates ] .full-width[ .conclusion[These principles have applications in material discovery, drug discovery, causal reasoning, etc. and have the potential of pushing the boundaries of machine learning research.] ]