search

SciML / DiffEqCallbacks.jl

A library of useful callbacks for hybrid scientific machine learning (SciML) with augmented differential equation solvers
https://docs.sciml.ai/DiffEqCallbacks/stable/
Other
95 stars 47 forks source link

Linearize derivatives at the same timepoints as primals #186

Closed staticfloat closed 9 months ago

staticfloat commented 9 months ago

After much discussion and investigation, it is decided that there is not an easy way to determine a good abstol/reltol to use for separate linearization of derivatives, and as such there is no reasonable rule of thumb for knowing how many timepoints is "enough" for the derivatives.

We instead make the same choice as the base solvers themselves; we sample the derivatives at precisely the timepoints of the primal. This yields an even more memory-efficient implementation where the derivatives can make use of the same ts and time_matrix as the primal.

staticfloat commented 9 months ago

@ChrisRackauckas This is ready to go from my perspective.