| title | menu_title | menu_icon |
|---|---|---|
Hackathon videos and slides |
Videos & slides |
film |
Below are talk recordings from the welcome and speaker session of the hackathon.
-
{% include youtube.html video="g4Pv3cVkSIg" title="Discovering materials twice as fast at a fraction of the cost through Bayesian optimization" %}
In this talk given at the BO Hackathon for Chemistry and Materials '24, University of Utah professor Dr. Taylor Sparks describes the need for good data to leverage the Acceleration that comes from using data-driven methods, and the role that Bayesian optimization plays in exploring materials search spaces. Dr. Sparks also addresses the concept of multi-objective optimization and the need to consider tradeoffs between multiple properties of interest through Pareto frontiers.
-
{% include youtube.html video="OKHcwtefRsU" title="Industrial view on Bayesian optimization: A perfect match for the low/no-data regime" %}
Dr. Martin Fitzner from Merck KGaA describes the role that Bayesian optimization plays in industrial chemistry and materials settings where experiments are expensive and data is often scarce. Dr. Fitzner describes the advantages of merging expert scientific intuition with sophisticated algorithms through the use of custom chemical encodings within a Bayesian optimization framework. His talk also addresses the potential to use transfer learning (i.e., multi-task BO) to kickstart the optimization of related tasks, and how their recently open-source Bayesian optimization BackEnd (BayBE) tool can provide a user-friendly and chemistry-focused experience with Bayesian optimization.
-
{% include youtube.html video="AIV_mvXpXIU" title="Bayesian Optimization for Sustainable Concrete" %}
Dr. Max Balandat from Meta's Adaptive Experimentation team describes the use of Bayesian optimization to reduce the carbon footprint of energy-intensive concrete formulations while retaining high performance properties through multi-objective optimization. The AI-suggested concrete mixes were experimentally verified and outperformed human-designed ones both in sustainability and strength. Dr. Balandat also described how domain knowledge could be incorporated into the search campaigns through the construction of the search space, such as use of log-scaling for a time-based parameter. Meta's open source Bayesian optimization tools, Ax and BoTorch, have been designed to address the diverse needs from researchers looking to create custom and cutting-edge solutions with modular building blocks (BoTorch) to a more plug-and-play and user-friendly interface (Ax). Finally, he points to the large potential of thoughtfully implementating Bayesian optimization for projects of impact.
This section will contain project presentations that were submitted by the teams.