2021-04-13 Shirts/Madin/Boothroyd Check in

Overall Framework

• SB: Hard problem. Big difficulty. All the software scientists are overworked, all the way to August. Thompson with system object, Biopolymer infrastructure.

• Not clear the software scientist have the background. If we SERIOUSLY want to have a force balance replacement. We would want a dedicated software scientist.

• What could we be short term w/o a full force balance replacement. Depends on the domain.

• If we just want to do the BCC with pyro or something like that.

• If we want to build surrogates with physical properties.

• We want something w/o having a big framework.

• Something less structured

• Can we start building force balance replacement for the QM side of things. More force balance

• SOME force balance replacement should go ahead.

• Get up a confluence page up with the scope and the outline.

• MRS: surrogate modeling.

• SB - In addition to work on Bayesian BCC refits or in place of (especially considering amount of time the OM has available)

• OM - BCC Bayesian fitting likely something we can do without too much infrastructure overhead. Probably something can do side by side but need to be careful.

• SB - Bayesian exploration of BCCs probably still quite impactful. Can help use to explore how much Bayesian optimisation will help us explore chemical perception issues, will we have sampling issues when sampling these higher dimensional, rough landscapes.

• SB: problem; everyone is stretched so thin. Expanding the scope slightly could push it over the edge.

• MRS: we want to bring in more resources.

• Showing we could sample with GPyTorch. Taking the alcohol dataset to start with.

• OM: starting on a dataset with can use a small number of measurements. Most of the parallelization/distribution is handled by evaluator. We can keep scaling until we hit a wall.

• SB: we can get all the derivatives easily. Might need to play with to see do we need to make the gradients.

• TODOS:

  •  

  • OM: