We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Anders Johansson - Fast and accurate machine learning potentials for extreme-scale simulations
Machine learning potentials for extreme-scale simulations: learn about the trade-offs between accuracy, speed, and performance portability, and discover promising tools like Allegro and Flare for creating accurate and scalable simulations.
- Machine learning is not always necessary: The speaker notes that some problems can be solved with guessing, rather than rigorous calculation.
- Accurate models can be created: Allegro models can be trained to be super accurate, and they’re compatible with PyTorch.
- Equivariant models are better: Equivariant models, such as Flare, tend to be more accurate than invariant models, and they respect symmetries.
- Uncertainties are important: Uncertainties in the model are crucial for accurate predictions, and Flare models include uncertainties.
- Performance portability is crucial: For simulation speed, performance portability becomes increasingly important, especially for large-scale simulations.
- GPU computing is essential: The speaker emphasizes the importance of using GPUs for machine learning simulations, particularly for molecular dynamics.
- Allegro and Flare are noteworthy: Allegro and Flare are machine learning potentials that are data-efficient, scalable, and accurate, making them promising tools for simulation.
- Slack channel is available: The speaker invites questions and comments on a Slack channel.
- LAMPs and COCOs are useful: LAMPs and COCOs are libraries that can aid in creating simulations with Allegro and Flare.
- More research is needed: The speaker suggests that more research is needed to further develop and refine machine learning potentials for simulation.