ExaChem: Exascale Electronic-Structure Package

• ExaChem is an open-source, exascale-oriented electronic-structure package that unifies quantum-chemistry and cavity QED methods using a common tensor framework.
• It leverages scalable algorithms and heterogeneous execution on GPU- and CPU-based systems to perform high-accuracy coupled-cluster and many-body simulations.
• The platform integrates advanced techniques like DUCC downfolding and DMRG to efficiently treat both dynamic and strong electron correlation in complex systems.

Searching arXiv for recent ExaChem-related papers to ground the article in published work. arxiv_search(query="ExaChem", max_results=10, sort_by="relevance") I’m going to look up ExaChem on arXiv so the article can cite the relevant papers directly. Searching arXiv for ExaChem-related records. ExaChem is an open-source, exascale-oriented electronic-structure package developed at PNNL that unifies standard quantum-chemistry methods with ab initio cavity quantum electrodynamics (QED) through a common tensor framework, and is also described as an open-source, high-performance platform for next-generation electronic-structure simulations (Pathak et al., 2024, Bauman et al., 15 Sep 2025). At its core is the Tensor Algebra for Many-body Methods (TAMM) infrastructure, which supplies the data abstraction and high-performance execution model needed to carry coupled-cluster, perturbation, and other many-body methods across GPU- and CPU-based leadership-class machines, workstations, and laptops (Pathak et al., 2024). In the current literature, ExaChem is presented along two closely related axes: as a scalable environment for conventional and QED-augmented coupled-cluster calculations, including QED-CCSD, and as a platform for high-accuracy workflows that combine coupled-cluster downfolding based on the Hermitian Double Unitary Coupled Cluster (DUCC) Ansatz with Density Matrix Renormalization Group (DMRG) solvers and hardware–software co-design involving CXL-attached memory and Azure Quantum Elements cloud resources (Pathak et al., 2024, Bauman et al., 15 Sep 2025).

1. Software identity and scientific scope

ExaChem’s stated objective is to make possible high-accuracy treatments of both static and dynamic correlation in molecular and materials systems—especially those exhibiting strong multireference character—by combining coupled-cluster downfolding based on the Hermitian Double Unitary Coupled Cluster (DUCC) Ansatz, Density Matrix Renormalization Group solvers in a matrix-product-state representation, and a hardware–software co-design across Micron CXL-attached memory pools for memory-heavy coupled-cluster steps and Microsoft Azure Quantum Elements cloud for DMRG sweeps (Bauman et al., 15 Sep 2025). In a separate but complementary direction, ExaChem extends conventional electronic coupled-cluster methodology to cavity QED by implementing QED-CCSD for electronic and bosonic amplitudes, including both individual and mixed excitation processes, within the same TAMM-based environment (Pathak et al., 2024).

This dual positioning is central to understanding ExaChem. One branch emphasizes standard many-body electronic-structure methods scaled to heterogeneous exascale systems; the other incorporates bosonic degrees of freedom and light–matter coupling into the coupled-cluster formalism. A further branch addresses strong correlation by compressing dynamical correlation into an effective Hamiltonian in a modestly sized active space and then solving the resulting problem by DMRG (Bauman et al., 15 Sep 2025). Taken together, these directions place ExaChem at the intersection of coupled-cluster theory, tensor libraries for heterogeneous HPC, cavity QED, and active-space many-body solvers.

A concise way to organize the package as described in the literature is given below.

Component	Role	Source context
TAMM	Tensor library and execution framework	Conventional CCSD and QED-CCSD
QED-CCSD	Electronic, bosonic, and mixed excitations	Cavity QED in the Pauli–Fierz framework
CCSD+DUCC module	Downfolding to effective Hamiltonians	Active-space reduction
MOLMPS	DMRG solver in MPS form	Strong-correlation solver phase

The package is therefore not a single-method code in a narrow sense. Rather, it is a common tensor and execution framework within which multiple high-accuracy many-body workflows are implemented.

2. TAMM infrastructure and heterogeneous execution model

At the heart of ExaChem lies TAMM, a parallel heterogeneous tensor library designed for utilizing modern computing platforms, from laptops to leadership-class computing resources (Pathak et al., 2024). Tensors in TAMM carry both index-label metadata and distribution directives, and each tensor is partitioned into blocks that can be placed on different MPI ranks or GPUs using a block-cyclic or user-specified pattern. TAMM supports both real- and complex-valued algebra, mixed precision (fp32/fp64), and nested tensor-of-tensor layouts to accommodate orbital-, spin-, and photon-mode indices (Pathak et al., 2024).

On distributed-memory clusters, TAMM uses Global Arrays for one-sided communication and MPI for collective operations. Within a node, highly optimized BLAS/LAPACK kernels handle CPU-resident blocks, while CUDA, or HIP/ROCm, kernels operate on GPU-resident blocks. To minimize CPU–GPU data transfers, TAMM employs “batched” contraction loops in which many small tensor blocks are summed locally on each GPU before any data is sent over the interconnect (Pathak et al., 2024). Task scheduling is driven by a DAG of tensor operations; each node in the DAG corresponds to a local contraction or transform, with dependencies enforced through MPI barriers only when necessary. Collective redistributions, described as tensor “transposes,” are implemented via pipelined one-sided puts/gets to hide communication behind computation (Pathak et al., 2024).

The same infrastructure underpins ExaChem’s conventional and QED workflows. On leadership-class machines, such as DOE supercomputers with hundreds to thousands of GPUs, ExaChem maps each tensor to as many MPI ranks as needed to saturate the network, achieving 70–80% parallel efficiency on up to 4096 ranks for conventional CCSD. On smaller servers or laptops, the same code paths automatically collapse to a single process, using multithreaded BLAS and, if present, a single GPU (Pathak et al., 2024).

In the DUCC/DMRG setting, TAMM is also identified as a C++ domain-specific library that generates highly optimized MPI+CUDA code for tensor contractions. SymGen performs automatic derivation and factorization of the relevant second-quantized expressions, and TAMM executes the resulting contractions (Bauman et al., 15 Sep 2025). This suggests that ExaChem’s central abstraction is not merely a set of chemistry kernels, but a tensor-programming substrate intended to support multiple many-body theories with a unified distribution and scheduling model.

3. QED-CCSD formulation within ExaChem

ExaChem’s QED-CCSD extends the usual electronic CCSD framework by including bosonic (photon) excitations and mixed electron–photon excitations (Pathak et al., 2024). The light–matter interaction is modeled using the Pauli–Fierz Hamiltonian within the dipole approximation in the length gauge. For a single cavity mode, the Hamiltonian is written in second-quantized form as

$$H = H_e + \Omega_c b^\dagger b - \sqrt{\Omega_c/2}\,\lambda\cdot(d-\langle d\rangle)(b^\dagger + b) + \frac{1}{2}\,[\lambda\cdot(d-\langle d\rangle)]^2.$$

Here,

$$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$

with $d$ the electronic dipole operator, $\Omega_c$ the cavity frequency, $\lambda$ the coupling-strength vector, and $b^\dagger/b$ the photon-creation/annihilation operators (Pathak et al., 2024).

The QED-CC wavefunction is built on a direct-product reference $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$ using the exponential ansatz

$$|\Psi_{\mathrm{QED\text{-}CC}}\rangle = e^T |R\rangle,$$

where the cluster operator includes pure electronic excitations $T_e$, pure bosonic excitations $T_b$, and mixed excitations $$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$0:

$$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$1

Truncating at single and double excitations in both sectors yields the QED-CCSD(2,2) model with the explicit forms

$$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$2

$$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$3

$$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$4

The stationary-condition equations are

$$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$5

for each combination of fermionic index $$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$6 and photon number $$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$7 up to two. The electronic-only subproblem with $$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$8 reduces to standard CCSD, while $$H_e = \sum_{pq} h_p^q\,E_q^p + \frac{1}{2}\sum_{pqrs} u_{pq}^{rs}\,E_{rs}^{pq},$$9 couples the photonic amplitudes (Pathak et al., 2024).

The significance of this implementation lies in the fact that bosonic and mixed electron–photon excitations are handled within the same tensor framework as conventional electronic amplitudes. This supports a comprehensive description of how cavity photons alter the electronic ground state and electron-correlation patterns.

4. Algorithms, tensor contractions, and convergence procedures

In ExaChem’s QED-CCSD implementation, all Hamiltonian-$d$0-$d$1 and $d$2-coupling diagrams are expressed as sequences of TAMM tensor contractions (Pathak et al., 2024). An illustrative mixed-excitation singles residual is reported as

$d$3

Each contraction is dispatched to the most efficient engine, described as GPU fused multiply-add or CPU GEMM. Electronic blocks, such as four-index two-electron integrals, and photon blocks, described as scalars or rank-one in $d$4, are distributed in block-cyclic fashion, and mixed electron–photon amplitudes inherit the electronic distribution. TAMM’s runtime scheduler builds a minimal-depth execution plan so that, wherever possible, a rank holds all operands locally for a contraction (Pathak et al., 2024).

The convergence procedure is specified at two levels. At the SCF level, for the QED-HF reference, the density-matrix residual is driven below $d$5 in density and $d$6 in Frobenius norm of the change, using DIIS with up to five vectors. In the coupled-cluster iterations, the maximum residual norm in each amplitude block—electronic singles and doubles together with photon-excited blocks—must fall below $d$7 before convergence is declared. A combination of DIIS and Jacobi-style mixing accelerates convergence (Pathak et al., 2024).

For the DUCC workflow, ExaChem’s operation is likewise split into a tensor-intensive construction stage and a solver stage (Bauman et al., 15 Sep 2025). In the downfolding phase, a standard CCSD calculation is carried out in the full orbital space $d$8 to extract external cluster amplitudes $d$9, assemble the DUCC(3) effective Hamiltonian $\Omega_c$0 in a chosen active space of size $\Omega_c$1, and generate one- and two-body tensor integrals $\Omega_c$2 and $\Omega_c$3. This step is both memory- and compute-intensive, scaling as $\Omega_c$4 for CCSD and $\Omega_c$5 in downfolding (Bauman et al., 15 Sep 2025).

The architecture of ExaChem’s algorithms is therefore characterized by an explicit reduction of formal many-body theory to tensor contractions, together with runtime decisions about locality, execution depth, and CPU/GPU placement.

5. DUCC downfolding, effective Hamiltonians, and DMRG

ExaChem’s coupled-cluster downfolding is based on the Hermitian DUCC Ansatz (Bauman et al., 15 Sep 2025). By splitting the cluster operator into “external” and “internal” parts, ExaChem constructs an effective Hamiltonian that acts only in the active-space projector $\Omega_c$6:

$\Omega_c$7

with

$\Omega_c$8

Here $\Omega_c$9 is the usual electronic Hamiltonian, $\lambda$0 is the Hartree–Fock projector, and $\lambda$1 projects onto excitations within the active space. Truncating the Baker–Campbell–Hausdorff series to include all double commutators and those triple commutators necessary for consistent CCSD-level accuracy yields DUCC(3). Rank-three and higher effective interactions are then discarded, so that the effective Hamiltonian retains only zero- through two-body terms:

$\lambda$2

In practice, the scalar $\lambda$3, one-body $\lambda$4, and two-body $\lambda$5 integrals are generated by summing more than a thousand Hugenholtz-type diagrams, whose automatic derivation and factorization is carried out by SymGen and TAMM (Bauman et al., 15 Sep 2025).

Once $\lambda$6 is available, ExaChem’s MOLMPS DMRG solver represents the active-space wavefunction as a Matrix Product State:

$\lambda$7

Here $\lambda$8 labels the four possible occupations of orbital $\lambda$9, and the virtual bond dimension $b^\dagger/b$0 controls the retained entanglement. MOLMPS implements parallel tensor-block distribution over MPI ranks, dynamical block-sizing based on the entanglement spectrum, efficient contraction kernels via TAMM, and an exposed API to ingest $b^\dagger/b$1 and $b^\dagger/b$2 from DUCC. Because $b^\dagger/b$3 is one- plus two-body only, no novel DMRG-specific integrals or three-body routines are needed; ExaChem simply replaces the bare integrals by the downfolded ones, and all DMRG machinery proceeds as usual (Bauman et al., 15 Sep 2025).

This workflow formalizes a particular division of labor between methods: coupled-cluster treats dynamical correlation in the full space, while DMRG resolves the active-space many-body problem in the effective Hamiltonian. A plausible implication is that ExaChem is designed not only for absolute scaling, but also for method orchestration across qualitatively different correlation regimes.

6. Performance characteristics, applications, and limitations

For QED-CCSD$b^\dagger/b$4, the computational cost remains $b^\dagger/b$5 in the number of orbitals, matching electronic CCSD, although mixed-excitation blocks enlarge prefactors (Pathak et al., 2024). On a 32-GPU partition of a DOE leadership machine, a water molecule in the cc-pVTZ basis with 30 occupied plus virtual orbitals converges in approximately 30 minutes to a $b^\dagger/b$6 amplitude residual, with approximately 75% strong scaling efficiency from 4 to 32 GPUs. Benchmarks reported for representative molecular systems include perfect agreement with QED-FCI on total energies and correlation energies to $b^\dagger/b$7 for $b^\dagger/b$8 from cc-pVDZ to cc-pVTZ under $b^\dagger/b$9 and $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$0 eV; seven-significant-figure agreement for water at experimental geometry from cc-pVDZ to aug-cc-pVTZ with $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$1 and $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$2 eV; and full QED-CCSD(2,2) calculations for malonaldehyde and aminopropenal completing in under one hour on 16 GPUs with modest, less than 10%, overhead relative to conventional CCSD (Pathak et al., 2024).

The reported bottlenecks depend on system size. For smaller basis sets, tensor redistribution, specifically the four-index integral transpose, dominates. For larger bases, the four-index–two-index contractions forming the doubles residual become rate-limiting. TAMM’s fused GEMM paths on GPUs are reported to mitigate both (Pathak et al., 2024).

In the DUCC/CXL/AQE context, ExaChem’s hallmark is explicit co-design of algorithm and hardware (Bauman et al., 15 Sep 2025). A Micron prototype cluster is described as using a single XCONN XC50256 CXL 2.0 switch feeding four Intel Xeon “Emerald Rapids” CPU hosts with 1 TB DRAM each and 15 FPGA–DIMM modules of 512 GB each, for a total of 7.5 TB shared memory, expandable to 33 TB per level or 132 TB with two-level switching. CXL memory modules are presented to the operating system in DAX mode, and ExaChem’s TAMM allocator is replaced by the Rapid library, which provides a software-coherent shared virtual address space across hosts. Typical CXL performance is reported as idle latency of approximately 520 ns for one-level switching and aggregate bandwidth of 264 GB/s per socket; in two-socket two-host runs, CCSD iteration times match or exceed GA-MPI versions when crossing host-boundary communications become limiting (Bauman et al., 15 Sep 2025).

Benchmarks reported in this setting include Fe(H$|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$3O)$|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$4 in cc-pVTZ with $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$5, where CCSD iteration time scales nearly linearly from approximately 8 s on GA-MPI with one host to approximately 6 s on CXL with two hosts. The DUCC(3) construction, involving approximately 1000 diagrams of double-commutator rank, completes in 30–45 minutes for $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$6 and $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$7 on four hosts. Memory usage is reported as approximately 500 GB for global CCSD and approximately 200 GB for DUCC(3) integrals in active indices, fitting within the 7.5 TB CXL pool with headroom for basis size up to approximately 500 or active-space size up to approximately 60 (Bauman et al., 15 Sep 2025).

For the DMRG solver phase on Azure Quantum Elements, ExaChem offloads sweeps to Slurm clusters supporting VM families including D8 v5, NV36 v5, HB120 v2, NC96 v4, and ND96 v5. Data movement from the CXL host to the cloud is limited to the $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$8 one- and two-body integrals, together with any DMRG checkpoint files. Reported applications include Fe–nitrosyl isomerization with CAS(8,8), (16,16), and (32,32) at bond dimensions $|R\rangle = |0^e\rangle \otimes |0^{ph}\rangle$9–2000, where single-sweep times range from 1 minute to 15 minutes on ND96 v5 nodes, and retinal cis–trans rotation with CAS(12,12) to CAS(26,26) and $$|\Psi_{\mathrm{QED\text{-}CC}}\rangle = e^T |R\rangle,$$0, where a full scan of approximately 19 angles takes approximately 5 h on an NV36 v5 partition. In the retinal study, DMRG-DUCC reproduces NEVPT2 energy profiles within 1 kcal/mol for small diradical index $$|\Psi_{\mathrm{QED\text{-}CC}}\rangle = e^T |R\rangle,$$1, diverging by up to 2.5 kcal/mol for highly open-shell regimes (Bauman et al., 15 Sep 2025).

Typical ExaChem use cases listed for the QED branch are ground-state energy shifts of molecules inside optical cavities, analysis of photon-induced correlation effects on proton-transfer or isomerization barriers, origin-invariance studies by shifting molecular coordinates and verifying invariance of QED-CCSD results, and systematic exploration of electron–photon coupling strength $$|\Psi_{\mathrm{QED\text{-}CC}}\rangle = e^T |R\rangle,$$2 and cavity frequency $$|\Psi_{\mathrm{QED\text{-}CC}}\rangle = e^T |R\rangle,$$3 on correlation energies and total energies (Pathak et al., 2024). Limitations explicitly noted in the DUCC/CXL/AQE work include the fact that CXL 2.0 only supports software coherence, that DUCC(3) limits the effective Hamiltonian to one- and two-body terms, that $$|\Psi_{\mathrm{QED\text{-}CC}}\rangle = e^T |R\rangle,$$4 is approximated by CCSD amplitudes, and that systems with strong three-electron interactions may require a rank-three effective Hamiltonian. Integration with emerging quantum hardware is stated to be on the development roadmap (Bauman et al., 15 Sep 2025, Bauman et al., 1 Jul 2025).

Taken together, these reports depict ExaChem as a tensor-centric electronic-structure environment in which conventional coupled-cluster theory, QED-augmented coupled-cluster theory, active-space downfolding, and DMRG are combined with explicit heterogeneous execution strategies. This suggests that the package’s main distinguishing feature is the attempt to treat method development, tensor algebra, and hardware placement as a single design problem rather than as separable layers.