Regular update #40
Merged
Regular update #40
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Test is currently failing; energy returned by excitations solver does not correspond to any eigenvalue of the lassi Hamiltonian so this is an error, not a local-minimum problem at the moment.
It was previously discarding the between-fragment part of the effective 1-electron operator, which is obviously not right
Self-consistency achieved, but it's stuck on local minima now.
Local minimum problem from previous commit fixed for 9 of the 11 perturber rootspaces.
Should not report failure to find global minimum in CI jobs, since I am still developing the algorithm.
Remove the unittest cases where it's numerically very difficult to avoid local minima, because they reflect unlikely use cases (direct spin excitation and double charge transfer).
set_excited_fragment_ now takes the value of the quantum numbers instead of the delta from the q-space, since there will be multiple q-spaces in the future. "excited" is now used in place of "active" throughout the code to clarify the relationship between the CAS and the space of excited fragments. docstring for set_excited_fragment_
Allows the Q space to have more than one rootspace in it, so that the stabilizing coupling of m,m' <-> m+1,m'-1 rootspaces is fully expressed in the vrvsolver.
Using the ExcitationSolver. Things that enabled this are 1) loosen self-consistency requirement of initial guess H(E)|Psi> = E|Psi> and 2) do not greedily solve H(E)|Psi> = E|Psi> for each fragment, but instead only do one pass of the undressed FCISolver kernel per VRVSolver kernel call. It is still very slow. As I learned in grad school, the fixed point algorithm for this fundamentally DOES NOT WORK when coupling is strong, at least as far as the energy determination is concerned. The good news is that H(E)|Psi> = E|Psi> is a strictly one-dimensional monotonic function between singularities so rootfinding it more properly should be easy.
Re-solve the self-consistent H(E)|Psi> = E|Psi> properly with fixed CI vectors on every cycle of the VRVSolver nonlinear iteration.
Update denom_q immediately upon computing e0; cleanup debugging stuff. There is still some failure to converge for alfefe nmax=0.
Slow calc'n compared to normal seems almost entirely due to repeated lassi op_o? function calls... is there any way to improve this?
- delete old commented code - e_p -> e0 consistently when I'm referring to the solution of self-consistent equations - the validity threshold of lowest_refovlp_eigval as a kwarg - some comment clarity
- _check_init_guess -> get_init_guess - sort_ci0 -> actually only sorts ci0 without also calling get_ham_pq
The VRV operators, H_PQ.(E-H_QQ)^-1.H_QP, will be averged over different P indices: sum_P w_P H_PQ.(E-H_QQ)^-1.H_QP
Doesn't reliably converge to anything, probably because I'm not quite understanding the equations I'm actually trying to solve when I do this. Needs theoretical development
in debug output in case of convergence failure
To permit the underlying _las object to be untouched by model-space manipulations.
Generates sz-rotated CI vectors directly from origin CI vectors.
Uses ExcitationPSFCISolver for charge-separated rootspaces, currently with equal-weights state-averaging. TODO: spin fluctations and unittests.
There is another warning for failure to converge anyway.
I can't believe I didn't have this before.
Copy the logic from lasci_sync.py over to lasscf_rdm.py
The lasscf_rdm kernel will now actually make sure that the coupled
fragment subproblems, which interact with each other via mean-
field ("jk") effects, have reached a fixed point before announcing
successful convergence. This is accomplished by a second micro-
iteration in which the orbitals are fixed by the fcibox kernels are
called repeatedly, taking the updated outputs of their neighbors
as inputs. Control the behavior of this new microiteration with
the "conv_tol_rdmjkddm", "conv_tol_rdmjkde", and "max_cycle_rdmjk"
attributes of the LASSCF method object. Also reduce default
"max_cycle_micro" to 3 from 5, since "max_cycle_micro" and
"max_cycle_rdmjk" handle separately what the lasci_sync kernel
handles simultaneously with "max_cycle_micro."
1. more conservative guardrail than last commit 2. transform_opmat_det2csf_pspace use integer range rather than bool array for indexing, since everything is supposed to be contiguous anyway 3. demote the eigenvalue test to even higher debug, since it becomes impossible fast
In case (esp in the excitation solver) it makes more sense to only optimize one fragment at a time.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
High-level summary: