Overview

EXESS (the Extreme-scale Electronic Structure System) is a GPU-accelerated quantum chemistry platform developed by QDX and made available through Rush. It is designed to deliver high-accuracy electronic structure calculations at scale, from high-throughput quantum workloads to chemically complex biomolecular environments. EXESS is built for cases where accuracy, throughput, and scalability must coexist.

A defining feature of EXESS is its regions-based workflow model. Regions provide a simple, explicit way to describe “what part of the system is treated with what level of theory,” enabling consistent execution across full-QM calculations, fragmentation workflows, QM/MM, QM/ML, and molecular dynamics—without rebuilding the workflow each time.

Regions: the organizing principle

In EXESS, regions are not just input syntax; they are the central abstraction that makes complex workflows practical and automatable. A regions specification lets you define, in a single place, how a system is partitioned into:

  • quantum regions (treated with HF (Hartree-Fock), KS-DFT (Kohn-Sham Density Functional Theory), RI-MP2 (Resolution of Identity Møller-Plesset second-order perturbation theory), CCSD (Coupled Cluster Singles and Doubles) energies, etc.)

  • environment regions (treated with classical MM (molecular mechanics) force fields and/or ML (machine learning) potentials)

  • fragment definitions and recombination logic for large heterogeneous systems

  • dynamic regions for simulations, where the chemically active region can be tracked and simulated over time

This structure is what allows EXESS to move smoothly between throughput-oriented quantum chemistry and large-system simulations, while keeping the workflow explicit, reproducible, and easy to automate.

High-throughput, multi-GPU quantum chemistry

EXESS provides a high-performance engine for non-fragmented calculations, including HF, KS-DFT, RI-MP2, and GPU-accelerated CCSD energies. These calculations are optimized around GPU-first computational kernels, multi-GPU scaling, and high throughput across large batches of independent jobs.

This supports workflows such as large-scale screening, geometry optimization, and reaction profiling, as well as the systematic generation of high-fidelity quantum mechanical reference data—energies, gradients, and electronic properties—that can be used to train, validate, and benchmark machine-learning models.

EXESS is not a machine-learning model generator. Its role is to generate accurate quantum data at a scale that makes ML training and validation genuinely data-driven rather than data-limited.

Fragmentation, QM/MM, and fragmentation-based ab initio molecular dynamics

For proteins and other large, heterogeneous systems, EXESS uses regions to express scientifically meaningful decompositions of the system and to make correlated QM methods tractable.

Fragmentation workflows partition a large system into chemically meaningful subsystems that are treated quantum mechanically and recombined to recover system-level properties at the target level of theory. These workflows underpin QDX’s largest biomolecular demonstrations with EXESS, particularly at MP2-class accuracy on GPUs.

Regions also provide a clean route to QM/MM (quantum mechanics/molecular mechanics) and hybrid embedding: chemically active regions (binding pockets, metal centers, reactive intermediates, covalent warheads) can be treated with high-accuracy QM, while the surrounding environment is treated with MM and/or ML. Because this is expressed through regions, the same setup naturally extends to fragmentation-based ab initio molecular dynamics, where long-time-scale simulations become practical by focusing high-level QM where it matters most and using scalable fragmentation/embedding for the remainder.

In drug discovery, this enables mechanistic simulations that capture polarization, charge transfer, bond formation and cleavage, and electronic reorganization in realistic biomolecular environments.

Integrated quantum–ML workflows on GPUs

EXESS includes native integration with machine-learning potentials within the same regions framework, enabling QM/ML hybrid workflows without breaking the execution model or leaving the GPU-first stack. These ML components are implemented from scratch with GPU performance as a primary design constraint.

The next major addition is NN-xTB (Neural Network extended Tight Binding), extending EXESS toward fast semi-empirical quantum models that retain a clear connection to electronic structure theory while benefiting from GPU acceleration and ML-driven parameterization. This complements high-accuracy QM by enabling efficient multi-fidelity workflows inside the same regions-driven pipeline.

Roadmap: pushing accuracy without losing scalability

EXESS is designed to evolve toward increasingly accurate quantum chemical methods while remaining GPU-native. Today, it supports CCSD energies on GPUs. Future development targets higher-level coupled-cluster methods (including perturbative triples corrections and beyond) and multi-reference approaches—quantum chemical methods designed for systems with strongly correlated electrons—tailored for chemically complex, heterogeneous systems.

All new methods follow the same guiding principle: they are designed natively for GPUs and multi-GPU execution, rather than adapted from CPU-first implementations.

Positioning

EXESS is built for workflows where quantum chemistry must be a production capability, not a bottleneck. Regions provide the glue that connects high-throughput quantum calculations, fragmentation at biomolecular scale, QM/MM and QM/ML embedding, and fragmentation-based ab initio molecular dynamics into a single coherent, automation-ready platform.

Rather than aiming for exhaustive feature coverage, EXESS emphasizes depth, performance, and scientific rigor—so that accurate quantum mechanics can scale both in system size and in workload volume.

What EXESS solves well

EXESS targets:

  • Large molecular systems where full-system QM is too expensive.

  • Fragment-based quantum chemistry via a Many-Body Expansion (MBE).

  • GPU-accelerated energies, gradients, geometry optimization, and AIMD.

  • Mixed-fidelity QM/MM/ML runs where fragments are assigned to regions (QMMM and optimization workflows).

  • High-throughput runs where a single input describes multiple topologies.

If your workflow needs plane waves or pseudopotentials, EXESS is not the right tool (see limitations). Thermostatted/barostatted dynamics are supported in QMMM workflows (NVT/NPT via qmmm.temperature_kelvin and qmmm.pressure_atm), but fully ab initio AIMD is microcanonical only.

Core ideas and workflow

EXESS is built around four practical ideas:

  1. GPU-first execution. Methods and kernels are designed to run efficiently on NVIDIA (CUDA) and AMD (HIP) GPUs.

  2. Fragmentation as a primary scaling strategy. Many-Body Expansion enables accurate calculations on systems that are otherwise too large for full-system QM.

  3. Region-aware modeling. The regions keyword assigns fragments to QM, MM, or ML partitions for QMMM and optimization workflows, with MM driven by OpenMM force fields and ML driven by AIMNet.

  4. Automation-friendly inputs. The input format supports batched topologies, which is useful for screening or dataset generation.

This philosophy shapes how you should approach method choice, fragmentation, and performance tuning.