JDC – Useful to have “flattened” versions of parameter assignment, and “non-flattened” versions
JDC – Helpful to look at API examples
MT – Jax needs a mathematical representation of how forcefield parameters map to system parameters
YTZ – Everything we do in Jax-land will be in terms of Q, not P.
JDC – System.to_jax will produce a
JDC – YTZ can compute dL/dQ, but it needs dL/dP
YTZ – I use “adjoint” in the opposite way that traditional practitioners do. I mean them to be dL/dQ.
(Recording started here)
JDC – As an exercise, maybe take timemachine’s harmonic bond (timemachine/potentials/bonded.py:30), and change it to take in a dL/dQ, and take in a dL/dP
JDC – Would be useful to look at API examples/points
YTZ – We should try to keep in mind that fitting to QM data will involve shipping different derivatives.
YTZ – “HW2”: Would be interesting to use jax.vjp to compute harmonic bond force. Would also be interesting to see dU/dTheta.
JDC – What about assigned parameters?