class LRDMCopt_workflow

This class manages the workflow of LRDMC optimization.

Module usage

from turboworkflows.workflow_lrdmcopt import LRDMCopt_workflow

Constructor arguments

job parameters

argument	type	default value	description
server_machine_name	str	localhost	name of the server machine for job submission.
queue_label	str	(None)	queue label for job submission.
mpi	bool	False	use MPI parallelization.
version	str	stable	version of packages to use.
sleep_time	int	1800	interval of job status check in seconds.

LRDMCopt parameters

argument	type	default value	description
lrdmcopt_max_continuation	int	2	maximum number of continuation runs.
lrdmcopt_target_error_bar	float	1.0e-3	target error bar in Ha.
lrdmcopt_trial_optsteps	int	50	number of trial optimization steps.
lrdmcopt_trial_steps	int	50	number of trial LRDMC steps per optimization step.
lrdmcopt_minimum_blocks	int	3	minimum number of blocks required.
lrdmcopt_production_optsteps	int	2000	number of production optimization steps.
lrdmcopt_optwarmupsteps_ratio	float	0.8	ratio of warmup steps to discard.
lrdmcopt_bin_block	int	1	binning length
lrdmcopt_warmupblocks	int	0	number of warmup blocks to discard.
lrdmcopt_optimizer	str	sr	optimizer type. options: sr for stochastic reconfiguration, lr for linear method.
lrdmcopt_learning_rate	float	0.002	optimization step size, default values is 0.05 for sr, 0.35 for lr.
lrdmcopt_regularization	float	0.001	regularization parameter.
lrdmcopt_num_opt_param	int	0	number of optimized parameters. 0 means all the parameters are optimized.
lrdmcopt_alat	float	-0.20	lattice spacing (Bohr)
lrdmcopt_time_branching	float	0.10	interval between two branching steps (a.u.)
lrdmcopt_trial_etry	float	0.0	trial energy (Ha)
lrdmcopt_nonlocalmoves	str	dla	treatment of locality approximation. options: tmove, dla, dlatm
lrdmcopt_onebody	bool	False	flag to optimize onebody Jastrow
lrdmcopt_twobody	bool	False	flag to optimize twobody Jastrow
lrdmcopt_det_mat	bool	True	flag to optimize matrix elements in the determinant part
lrdmcopt_jas_mat	bool	False	flag to optimize matrix elements in the Jastrow part
lrdmcopt_det_basis_exp	bool	False	flag to optimize exponents of the determinant basis sets
lrdmcopt_jas_basis_exp	bool	False	flag to optimize exponents of the Jastrow basis sets
lrdmcopt_det_basis_coeff	bool	False	flag to optimize coefficients of the determinant basis sets
lrdmcopt_jas_basis_coeff	bool	False	flag to optimize coefficients of the Jastrow basis sets
lrdmcopt_num_walkers	int	-1	number of walkers, -1 (default) = the number of MPI processes
lrdmcopt_twist_average	bool	False	flag for twist average
lrdmcopt_kpoints	list	[]	Monkhorst-Pack k-grids, [kx,ky,kz,nx,ny,nz], kx,ky,kz for grids, nx,ny,nz for shift=0, noshift=1.
lrdmcopt_maxtime	int	172000	maximum time (sec.)

Description

In the LRDMCopt_workflow class, the LRDMC optimization workflow is executed asynchronously (via async_launch()). The workflow runs LRDMC optimization in a continuation loop (up to lrdmcopt_max_continuation runs). For each continuation index \(i_{\mathrm{cont}}\):

test run (icont = 0)

Uses fixed trial parameters (lrdmcopt_trial_optsteps, lrdmcopt_trial_steps) to produce energy and error-bar data.
No step estimation is performed.
Results are used by the next run to estimate required LRDMC steps.

production run (icont ≥ 1)

Loads the previous run's results from the corresponding pkl file.
Estimates the LRDMC steps required to reach lrdmcopt_target_error_bar via \(N_{\mathrm{steps}} \propto (\sigma / E_{\mathrm{target}})^2\).
Applies a minimum-step floor using lrdmcopt_minimum_blocks and lrdmcopt_warmupblocks.
Estimates total runtime and builds a LRDMCopt_genius instance with the estimated steps and fixed lrdmcopt_production_optsteps, then generates input and submits the job.

For each run, the workflow waits for job submission and completion (using asyncio.sleep()), fetches output files (e.g. fort.10, fort.11, fort.12, out_fn_opt_*, parminimized.d), stores and plots results, and for \(i_{\mathrm{cont}} > 0\) computes parameter averages discarding an initial fraction (lrdmcopt_optwarmupsteps_ratio). State is persisted in pkl files under the pkl directory.

If all continuation pkl files already exist and lrdmcopt_rerun is False, the calculation is skipped. On success, the method returns (status, list of output file paths, and an empty output-values dict).