DMRG-DUCC Workflow: Hybrid Downfolding & DMRG

• DMRG-DUCC is a hybrid active-space method that combines DUCC downfolding to integrate dynamic correlation with DMRG’s nonperturbative treatment of static correlation.
• The workflow constructs an effective Hamiltonian via a truncated Baker–Campbell–Hausdorff expansion, incorporating inactive orbital effects into active-space integrals.
• DMRG-DUCC improves energy convergence and entanglement representation in systems like N₂, benzene, and TME, balancing computational efficiency with accuracy.

Searching arXiv for the most relevant DMRG-DUCC papers and related implementation/context papers. The DMRG-DUCC workflow is a hybrid active-space electronic-structure strategy in which a Hermitian double-unitary coupled-cluster (DUCC) downfolding step first incorporates external, predominantly dynamical correlation into an effective Hamiltonian, and density matrix renormalization group (DMRG) then solves that renormalized Hamiltonian nonperturbatively within the chosen active space. In this formulation, DMRG is not applied to the bare active-space Hamiltonian, but to a dressed Hamiltonian that already contains much of the influence of the inactive orbitals. The workflow is therefore a canonical example of a “downfold-then-solve” architecture: coupled-cluster theory supplies an effective model, and DMRG supplies the multireference solver (Bauman et al., 2024).

1. Conceptual structure and scope

DMRG-DUCC is motivated by a separation of correlation regimes. Static correlation is associated with near-degenerate configurations inside an active space and must be treated nonperturbatively; DMRG is well suited to that regime. Dynamic correlation is associated with numerous weaker excitations involving orbitals outside the active space and is usually handled efficiently by single-reference coupled-cluster methods. The DMRG-DUCC workflow assigns these two tasks to different stages: DUCC folds the external-space effects into an active-space Hamiltonian, and DMRG then diagonalizes that Hamiltonian (Bauman et al., 2024).

This division of labor distinguishes the method from several neighboring approaches. Compared with conventional DMRG, the distinction is the Hamiltonian: ordinary active-space DMRG works with the bare projected Hamiltonian and therefore omits correlation effects associated with excluded orbitals. Compared with standard active-space methods such as CASCI and CASSCF, DMRG-DUCC does not merely solve the active-space problem more accurately; it changes the active-space Hamiltonian itself by incorporating inactive-space contributions. Compared with ordinary coupled cluster, CC is not the terminal solver; it is used only to generate the effective Hamiltonian, while the final many-body problem is solved by DMRG (Bauman et al., 2024).

The active-space choices reported for the method reflect this philosophy. For N$_2$, the reported calculations used CAS(14,13), CAS(14,19), and CAS(14,23) in cc-pVDZ. For benzene, the reported active spaces were CAS(42,32), CAS(42,36), CAS(42,39), and CAS(42,52). For tetramethyleneethane (TME), the reported active space was CAS(44,34). These choices place the expected multireference physics inside the DMRG domain and leave the predominantly dynamical remainder to DUCC (Bauman et al., 2024).

2. DUCC downfolding and the effective Hamiltonian

The starting point is the standard second-quantized electronic Hamiltonian,

$$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$

DUCC defines a Hermitian active-space effective Hamiltonian

$$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$

where $P$ projects onto the reference determinant $|\Phi\rangle$, typically Hartree–Fock, and $Q_{\rm int}$ projects onto excitations that remain entirely within the complete active space. The anti-Hermitian external operator is approximated as

$$\sigma_{\rm ext} \simeq T_{\rm ext}-T_{\rm ext}^\dagger,$$

with $T_{\rm ext}$ taken from the external part of the CCSD cluster operator. In the reported implementation, $T_{\rm ext}$ is therefore approximated by CCSD amplitudes involving at least one inactive spin-orbital index (Bauman et al., 2024).

Because the exact similarity transformation is non-terminating, the workflow uses a truncated Baker–Campbell–Hausdorff expansion. Two levels are reported in the 2024 study: DUCC(2), consistent through second-order perturbation theory, and DUCC(3), consistent through third-order perturbation theory (Bauman et al., 2024). In the later integrated execution model, the production DUCC construction is described more specifically as including all double commutators plus triple-commutator terms stemming from the Fock operator for perturbative consistency, followed by truncation to scalar, one-body, and two-body terms only (Bauman et al., 15 Sep 2025).

After truncation, the effective Hamiltonian takes the form

$$H^{\rm eff}\simeq \Gamma_0 + \sum_{pq} g^p_q\, a^\dagger_p a_q +\frac{1}{4}\sum_{pqrs} k^{pq}_{rs}\, a^\dagger_p a^\dagger_q a_s a_r,$$

where $$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$0 is a scalar shift and $$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$1, $$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$2 are effective one- and two-body tensors defined only over active spin-orbitals. The operative interpretation is that the external space has been integrated out into modified active-space integrals plus a constant term. The remaining active-space problem is therefore not the bare CAS Hamiltonian but a renormalized Hamiltonian that already encodes much of the missing dynamical correlation (Bauman et al., 2024).

A notable implementation point in the later work is that the updated DUCC implementation allows arbitrary active spaces, whereas earlier DUCC work had assumed that the active space contained all occupied orbitals. The same implementation also uses the frozen-core approximation, with occupied orbitals outside the active space treated as frozen (Bauman et al., 15 Sep 2025).

3. DMRG solution stage and numerical realization

Once $$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$3 has been constructed, it is passed to DMRG as an active-space Hamiltonian in standard second-quantized form. The DMRG stage solves

$$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$4

by representing the active-space wavefunction as a matrix product state (MPS) with bond dimension $$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$5, and optimizing the MPS by sweeps along the orbital chain (Bauman et al., 2024).

The reported implementation uses a two-site DMRG algorithm. At each step, a neighboring pair of orbitals is optimized while the remainder of the chain is encoded in left and right renormalized blocks. The current local effective Hamiltonian is diagonalized, the local tensor is updated, and the environment blocks are renormalized. The bond dimension controls accuracy through reduced-density-matrix truncation, so $$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$6 is the principal DMRG convergence parameter (Bauman et al., 2024). In a broader quantum-chemical DMRG context, the accuracy is not governed by $$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$7 alone but by the joint choice of active space, orbital basis, orbital ordering, initial guess, sweep strategy, and extrapolation protocol. The DMRG accuracy study emphasizes that the active space is the foundational approximation, that two-site DMRG is more robust than single-site DMRG, and that truncation-error extrapolation is a key diagnostic for quantitative work (Keller et al., 2014).

The active-space orbitals used in the DMRG-DUCC benchmarks were usually RHF orbitals, with additional RHF-versus-CCSD-natural-orbital comparisons for N$$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$8. The reported recommendation is that RHF orbitals are a more unbiased starting basis for general DUCC downfolding, while CCSD natural orbitals can help near equilibrium when the underlying CC description is reliable but become less robust at stretched geometries (Bauman et al., 2024).

The paper also remarks that DUCC downfolding tends to “compress” correlation into the active space. In the reported analyses of natural occupations, single-orbital entropies, and mutual information for N$$H_{\mathrm{el}}= \sum_{\sigma}\sum_{pq} h_{pq}\, a^\dagger_{p_\sigma} a_{q_\sigma} +\frac{1}{2}\sum_{\sigma\sigma'}\sum_{pqrs}\langle pq|rs\rangle a^\dagger_{p_\sigma} a^\dagger_{q_{\sigma'}} a_{s_{\sigma'}} a_{r_\sigma}.$$9, DUCC downfolding increased occupations of occupied orbitals and reduced virtual occupations, yielding a more closed-shell, less multireference active-space picture while redistributing correlation across the active manifold (Bauman et al., 2024). This suggests that the effective Hamiltonian is not merely shifting the total energy, but also altering the entanglement structure presented to DMRG.

4. Benchmark behavior and reported performance

The reported molecular benchmarks are designed to test different balances of static and dynamic correlation. N$$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$0 probes the crossover from equilibrium single-reference behavior to stretched-bond multireference behavior. Benzene probes a weak-static-correlation but still substantially correlated regime. TME probes a more challenging multireference potential-energy profile. Later work extends the reported tests to Fe–nitrosyl isomerization and retinal cis–trans isomerization (Bauman et al., 2024).

System	Setup	Reported outcome
N$$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$1	CAS(14,13), CAS(14,19), CAS(14,23) in cc-pVDZ	In CAS(14,13) with RHF orbitals at $$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$2, bare DMRG error is 221.1 mH; DMRG-DUCC(3) reduces it to 7.5 mH
Benzene	CAS(42,32), CAS(42,36), CAS(42,39), CAS(42,52)	DMRG-DUCC(2), especially DMRG-DUCC(3), gives substantially lower energies than bare DMRG; DUCC(3) energies lie between CCSD and CCSD(T)
TME	CAS(44,34)	Ordinary DMRG underestimates correlation energy and misorders relative energies; DMRG-DUCC places the 45-degree structure above the 0- and 90-degree structures
Fe–nitrosyl isomers	CAS(8,8), CAS(16,16), CAS(32,32)	Bare DMRG shows inconsistent gaps; DMRG-DUCC yields more systematic convergence, flat–standard gap around 3.47 eV, reverse–standard gap decreasing from 1.25 eV to 0.77 eV
Retinal cis–trans isomerization	cc-pVDZ, 386 orbitals, extended CAS(26,26)	DMRG-DUCC improves over CCSD but underestimates static correlation near $$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$3

For N$$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$4, the reported conclusion is that DMRG-DUCC dramatically improves equilibrium energies in small active spaces and continues to improve the stretched-bond regime even when bare active-space DMRG struggles (Bauman et al., 2024). For benzene, the central observation is stability with respect to active-space size: the DUCC-dressed Hamiltonian yields substantially lower energies than bare DMRG across the tested active spaces, and DUCC(3) lies between CCSD and CCSD(T), which the authors use as evidence that the dressed-Hamiltonian treatment recovers effects beyond the nominal CCSD input (Bauman et al., 2024). For TME, the method changes not only absolute energies but also the shape of the torsional profile, with reported twisting barriers of about 0.35 eV for DUCC(2) and about 0.4 eV for DUCC(3) (Bauman et al., 2024).

The later execution-model paper broadens the empirical picture. For Fe–nitrosyl isomerization, DMRG-DUCC yields more systematic active-space convergence than bare DMRG and larger gap magnitudes, which is attributed to restoration of dynamical correlation (Bauman et al., 15 Sep 2025). For retinal, the same study reports a more restrictive outcome: CASSCF, DMRG, and NEVPT2 yield smooth profiles, CCSD fails near the twisted geometry because of strong static correlation, and DMRG-DUCC improves over CCSD but still underestimates static correlation near the diradical point because the DUCC amplitudes remain CCSD-based. Near $$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$5, the reported occupations are $$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$6, $$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$7 for bare DMRG, but roughly $$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$8, $$H^{\rm eff}=(P+Q_{\rm int})\, e^{-\sigma_{\rm ext}} H e^{\sigma_{\rm ext}}\, (P+Q_{\rm int}),$$9 for DMRG-DUCC (Bauman et al., 15 Sep 2025).

5. Software stack, data flow, and heterogeneous execution

The 2024 implementation constructs the effective Hamiltonian in HDCC and passes the resulting tensors to the MOLMPS DMRG solver. The DMRG engine supports distributed matrix operations on CPUs and GPUs, while the effective integrals $P$0 and $P$1 are communicated through TAMM. The paper also notes that the lower symmetry of the DUCC-dressed tensors, relative to ordinary molecular-orbital integrals, reduces exploitable permutational symmetry in storage (Bauman et al., 2024).

The 2025 execution-model paper makes the hardware partition explicit. The workflow is decomposed into a CCSD/DUCC downfolding phase and a DMRG phase because the two stages stress different resources. The CCSD stage scales as $P$2, and the CC downfolding stage is reported to scale as $P$3, where $P$4 is the number of active orbitals and $P$5 is the total system size. These stages are characterized as memory- and tensor-contraction-intensive. The DMRG phase is characterized as less memory-intensive per sweep but still sensitive to bond dimension and active-space size, with rough per-sweep scaling

$P$6

On that basis, the reported integrated execution model runs CCSD and DUCC on Micron CXL 2.0 memory hardware and DMRG on Azure Quantum Elements cloud infrastructure (Bauman et al., 15 Sep 2025).

The same study reports that the DUCC(3) Hamiltonian contains over a thousand Hugenholtz-type diagrams, derived automatically with SymGen, translated to tensor expressions, converted to TAMM format, and executed in ExaChem. To exploit fabric-attached memory, TAMM allocation was modified so that tensors are allocated by rank 0 through the Rapid allocator, the pointer is broadcast, and tensors become directly accessible by all ranks without temporary buffers or extra copies. The prototype system lacks hardware multi-host coherency, so a software cache-coherency model with explicit flush/fence sequences is used (Bauman et al., 15 Sep 2025).

Within the broader DMRG software ecosystem, Block2 is relevant as a tensor-network backend rather than as a DUCC implementation. The Block2 framework directly provides ab initio DMRG, MPS/MPO infrastructure, symmetry-adapted calculations, excited-state algorithms, response methods, finite-temperature routines, density matrices, transition density matrices, and symbolic support for similarity-transformed Hamiltonians, but the paper on Block2 does not explicitly discuss DUCC itself (Zhai et al., 2023). A plausible implication is that DUCC would sit as an external downfolding layer around such a DMRG/MPS engine rather than as a native Block2 module.

6. Accuracy domain, failure modes, and relation to adjacent workflows

The main limitations reported for DMRG-DUCC are inherited from the quality of the external CC description and from truncations in the downfolding. The workflow relies on $P$7 from CCSD and on a truncated BCH expansion, so its quality degrades when CCSD becomes unreliable. The 2024 study states that the method works best when the active space captures the essential static correlation, the external space is dominated by dynamic correlation, CCSD is at least a reasonable approximation for the external amplitudes, and the DMRG calculation is sufficiently flexible (Bauman et al., 2024). The 2025 study sharpens this by introducing a diradical coefficient

$P$8

and reporting that when $P$9, DMRG-DUCC and NEVPT2 energy differences differ by at most about 2.5 kcal/mol, whereas for larger $|\Phi\rangle$0 the method increasingly underestimates static correlation (Bauman et al., 15 Sep 2025).

These limitations make the DMRG stage itself diagnostically important. In chemical DMRG more generally, the active space, orbital basis, orbital ordering, initial guess, bond dimension $|\Phi\rangle$1, sweep convergence, and truncation-error extrapolation all materially affect accuracy. The broader DMRG accuracy analysis explicitly warns that even a perfectly converged DMRG wavefunction remains chemically incomplete if the active space is too small, and that single-site DMRG can become trapped in local minima. For a DMRG-DUCC workflow, this means that DUCC is only as reliable as the DMRG reference problem and active-space definition allow (Keller et al., 2014).

DMRG-DUCC also occupies a distinct position relative to other DMRG-based dynamic-correlation workflows. In the DMRG-ec-MRCISD(+Q) procedure, the DMRG wavefunction is first converted into a truncated CASCI-type determinant expansion via the entropy-driving genetic algorithm (EDGA), and that determinant set is then used as a reference for externally contracted MRCISD with a Davidson correction. That approach avoids the high-order reduced-density-matrix bottleneck of internally contracted DMRG-MRPT and DMRG-MRCI by reconstructing only a small determinant subset and contracting the excited space externally (Luo et al., 2018). DMRG-DUCC, by contrast, modifies the Hamiltonian before the DMRG solve rather than reconstructing a post-DMRG determinant reference. The two workflows therefore solve different bottlenecks: DMRG-ec-MRCISD(+Q) is a post-reference dynamic-correlation correction built on selected determinants, whereas DMRG-DUCC is a pre-solve downfolding strategy built on a Hermitian active-space effective Hamiltonian.

In this sense, DMRG-DUCC is best understood as a hybrid multireference solver architecture rather than merely a DMRG correction. Its defining feature is the relocation of external correlation from the wavefunction ansatz into the Hamiltonian itself. That design is what allows a comparatively modest active-space DMRG calculation to recover a substantially larger portion of the total correlation energy than bare active-space DMRG, while retaining DMRG’s ability to treat strong static correlation nonperturbatively (Bauman et al., 2024).