DMRG-CASSCF

For ordinary YAML input, use the CASCI/CASSCF solver adapters. This page describes the separate manual Python route: use Block inside PySCF CASSCF and export the completed object through the external-reference API. Its maintained example is test_suite/workflows/external_reference/h4_dmrg_casscf/.

PySCF RHF → CASSCF orbital optimization with Block DMRG
          → reference_from_casscf → HDF5 → CIPSIXX → DIAGPT → HEFFSO

Install The Optional Solver

Complete the MPPT installation first.

Alternatively, create the separate mppt-block2 environment with block2 0.5.4 and the same pinned adapter. That environment already contains the optional solver packages; omit the pip install command below to retain its pins. The older manual requirements and measured baseline below use 0.5.3.

From the example folder:

python -m pip install -r requirements-manual.txt
OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 python run.py

The optional requirements pin Block2 0.5.3 and a known PySCF dmrgscf revision. They are not part of MPPT’s default environment. block2main and all three MPPT executables must be on PATH. BLOCKEXE=/absolute/path/to/block.spin_adapted selects an installed StackBlock instead; that backend has not been validated with this example. The runner uses one MPI rank and explicit OpenMP threads.

Replace The Active-Space Solver

The core change is small:

from pyscf import dmrgscf, mcscf

# Keep PySCF orbital optimization and replace only the active-space solver.
mc = mcscf.CASSCF(mf, ncas, nelecas)
mc.fcisolver = dmrgscf.DMRGCI(mol, maxM=100, tol=1.0e-10)

# Isolate Block's files and set its resource controls in the expected units.
mc.fcisolver.runtimeDir = str(block_workdir)
mc.fcisolver.scratchDirectory = str(block_workdir)
mc.fcisolver.threads = 1
mc.fcisolver.memory = 2  # GB, unlike PySCF max_memory in MB.

# Complete the orbital optimization before exporting this reference to MPPT.
mc.kernel()

This fragment assumes a configured dmrgscf installation and a converged mf. The executable example’s block_setup.py configures the supported PySCF hooks before importing the optional module, avoiding edits to site-packages.

Call reference_from_casscf(mc) after convergence and pass the returned object to dump_reference(). The external-reference guide documents final-Hamiltonian, selected-space and property-integral requirements.

Resume With The Same Workflow YAML

# Produce the reference first; this command stops before the MPPT stages.
python run.py --dump-only --output-dir runs/export-only

# Use the same workflow YAML to continue from that HDF5 via the CLI.
cd runs/export-only
pyscf2mppt ../../workflow.yaml --from-hdf5 reference.h5

run.py performs the same continuation by default. New SCF/export runs require fresh output directories. Keep HDF5 and downstream sidecar outputs together. CLI continuation creates runs/workflow/ below the export directory; repeat the CLI command with --resume to reuse that run. Native binary results live under data/, raw logs under logs/, and analysis tables under reports/heffso/.

What The Example Verifies

H4/STO-3G uses CASSCF(4,4), a Block bond dimension of 100 and a full 36-determinant CIPSIXX initial CAS. It compares DMRG-CASSCF with independent PySCF FCI within 1e-7 Hartree, and scalar/final MPPT energies within 1e-8 Hartree.

The validated DMRG-CASSCF energy is -2.16638744863368 Hartree with PySCF 2.14.0 and Block2 0.5.3; its FCI difference is 1.083e-12 Hartree. This is a small integration check, not a DMRG scaling benchmark or a nonzero-SOC example.

CTest marks the case Manual;ExternalReference;Block;DMRG. Missing optional prerequisites produce a skip; numerical or solver failures remain failures.

Large Active Spaces

MPPT consumes the optimized orbitals and Hamiltonian, not the Block MPS or its CI coefficients. CIPSIXX runs an independent selection. Its model space must be chosen explicitly; a large DMRG CAS must not be blindly expanded into all determinants. Choose bounded selection and validate the intended states.

This route retains the current restricted, real-spatial-orbital and C1 contract. DMRG optimization does not remove integral-export or downstream scaling limits. Use compact selected-route export when a full two-electron table is too large.