Skip to content
zyliu-OUPublic

About

Under development: experimental CPU/GPU GKS-EDA (six components, counterpoise, RKS/UKS, M06-2X) with DFT gradients and geometry optimization. Energies, gradients and minima cross-checked against Psi4; tryptamine–water totals match the 2017 paper within 0.012 kcal/mol. Matched XEDA component validation pending.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

EDA_GPU

Status: under active development — an early research prototype.

EDA_GPU is a GPU-assisted computational chemistry project for reproducible molecular structure preparation, electronic-energy calculations, analytic gradients, and geometry optimization. Its long-term goal is a faithful implementation of the original generalized Kohn–Sham energy decomposition analysis (GKS-EDA), beginning with a validated CPU reference and then accelerating the expensive operations on NVIDIA GPUs.

The working foundation uses PySCF for CPU calculations, GPU4PySCF for GPU calculations, and geomeTRIC for optimization. Current validation covers small, restricted closed-shell and small unrestricted gas-phase systems. APIs, supported methods, and installation requirements may change as development proceeds.

See scientific references, research papers, and websites for method attribution, benchmark sources, and software documentation.

What works today

Capability Current status
Explicit CPU/GPU backend selection Implemented and tested; unavailable GPU requests fail clearly
XYZ plus identity, protonation, charge, spin, and fragment metadata Implemented with input consistency checks
DFT single points and analytic nuclear gradients Tested with density-fitted B3LYP/def2-SVP
Water and hydrogen-bonded water-dimer optimization Convergence demonstrated on CPU and GPU
Frozen-atom optimization Tested on CPU and GPU; external freeze/set constraints also accepted
Machine-readable calculation results Settings, units, convergence, versions, hardware, timing, and chemical metadata
Optimized structures and trajectories XYZ outputs and evaluated-geometry trajectories
Scientific validation and CPU CI CPU/GPU agreement, finite differences, constraints, and explicit failure handling
Independent cross-program check HF/DFT energies and analytic gradients (RKS, UKS, exact and density-fitted) agree with Psi4 1.11 on CPU and GPU (21/21 within preset tolerances, up to 425 basis functions); optimized minima agree for water and water dimer; not an EDA validation
Published tryptamine–water benchmark Neutral and radical-cation totals within 0.012 kcal/mol of Liu et al. 2017 Table 3, components within 0.20 kcal/mol, H-bond geometry within 0.05 Å; reconstructed geometry. Accepted by the project owner (2026-10-19) as the external component benchmark
Matched XEDA comparison GAMESS-XEDA inputs and EDA_GPU values prepared (reports/xeda_matched_2026-10-18/); not run — deferred by project decision 2026-10-19 (kit kept for anyone with XEDA access)
Experimental GKS-EDA(sol) C-PCM/IEF-PCM on CPU and GPU; vacuum-limit, screening and CPU/GPU checks pass; documented interpretations; external validation pending
CPU/GPU scaling benchmark GPU 3–7× faster than CPU from ~250 basis functions on this host; slower for tiny systems (benchmark)
CPU EDA preparation Counterpoise fragment references, frozen electrostatics, generalized correlation; independent algebra checks passed, published component validation unresolved
Complete experimental GKS-EDA Six gas-phase terms on CPU/GPU; independent mathematical tests and component-wise CPU/GPU agreement passed; external matched-settings validation pending

The CLI provides inspect, singlepoint, optimize, validate, and experimental eda-prepare and eda commands. This project implements an independently useful optimization foundation; it does not claim a GAMESS port or a validated EDA decomposition.

Validation evidence

Test computer and software environment

The local tests and neutral tryptamine–water benchmark were run on the following computer. These are the recorded test specifications, not minimum requirements.

Item Test environment
CPU AMD Ryzen 9 5950X; 16 physical cores, 32 logical threads
GPU NVIDIA GeForce RTX 4080 SUPER; 16 GB VRAM (16,376 MiB reported)
System memory 30.25 GiB usable RAM reported by Linux
Operating system Ubuntu 26.04 LTS, x86-64
NVIDIA driver 595.91.07
Python 3.12.13
PySCF / GPU4PySCF 2.14.0 / 1.8.1
CuPy / cuTensor 14.2.0 / 2.8.1
geomeTRIC 1.1.1

The archived neutral EDA run used 16 physical cores/compute threads, 50% workload pacing, and CPU/GPU thermal cooldown thresholds of 85°C/75°C. Earlier small-system timings used their own recorded settings and are not directly comparable with this paced run. GPU power limits and temporary CPU frequency limits are shared hardware settings; no fixed frequency or power setting is claimed for all tests. Exact per-run hardware, versions, and calculation settings are saved in result JSON.

On the development host (RTX 4080 SUPER, Ryzen 9 5950X, Ubuntu 26.04), 37 local tests passed in the latest full suite, including actual GPU tests. Separately, nine optimization-foundation scientific checks passed with actual GPU execution. Water and water-dimer CPU/GPU single-point energies differed by at most 9.21e-12 hartree; maximum gradient differences were 7.35e-7 hartree/bohr. Water finite-difference gradients agreed within 3.59e-6 hartree/bohr, and frozen oxygen displacement was below 1.79e-11 angstrom. Both molecules converged on both backends.

These results apply to the recorded settings and small benchmark cases. Optimization convergence establishes a stationary-point result; a vibrational analysis has not established a minimum. GitHub CPU CI also passed; its GPU test was explicitly skipped because that runner has no GPU. Warmed, repeated timing measurements are provided without extrapolating them to larger systems.

See validation details, full scientific results, and benchmark measurements.

Development roadmap

  1. Extend the structure-preparation and optimization foundation with additional carefully validated molecules, installation paths, and supported settings.
  2. Secure the original GKS-EDA derivation, supporting material, and trustworthy individual component reference calculations under identical settings.
  3. Specify every intermediate state and component convention, including orbital normalization, exchange/correlation, fragment reference geometry, and BSSE policy (experimental closed-shell specification now implemented).
  4. Validate the implemented experimental CPU GKS-EDA against external matched-settings outputs, component by component.
  5. Extend the implemented GPU component path after external validation; improve performance beyond the current small-system direct-integral measurements.
  6. Nucleotide/RNA–ion-pair reagent–stationary-phase models: first results and a comparison with published retention data are in the HPLC application study; a nonpolar solvation term and placement sampling are needed before retention trends can be interpreted.

Complete restricted/unrestricted gas-phase EDA reproduces published GKS-EDA values to ≤0.20 kcal/mol per component (see the benchmark row above); GKS-EDA(sol) and the HPLC study remain experimental. Total-energy closure alone will not count as component validation, and ALMO-EDA or other decompositions will not be substituted for the original GKS-EDA definitions. The exact current blockers are documented in the GKS-EDA source gate.

HPLC application study (experimental)

An application of GKS-EDA(sol) to ion-pair reversed-phase (IP-RP) HPLC of RNA: how a capped RNA residue (nucleotide with a −1 phosphodiester), an alkylammonium ion-pair reagent and a C8 stationary-phase proxy (octyldimethylsilyl–O–Si(OH)₃) interact. All structures were optimized on the GPU at M06-2X/def2-SVP in C-PCM water (frozen silane anchor), then decomposed with three-body GKS-EDA(sol). This is exploratory work: one starting geometry per case, constrained stationary points, no cavitation term in the solvent model. Full tables: validation; models and limits: HPLC models; figures: scripts/plot_hplc_figures.py.

Optimized model systems

Model systems compared (archives under reports/):

System What it models Cases Archive
Ion pair (residue···reagent) phosphate–ammonium salt bridge in solution 4 bases × TEAH⁺; residue A × 5 reagents hplc_ionpair_2026-10-19/, hplc_reagents_2026-10-21/
Residue···C8 / reagent···C8 direct adsorption on the proxy 4 bases; 5 reagents (17 placements) hplc_surface_2026-10-19/, hplc_reagents_2026-10-21/
Silanol-bound ternary residue on a silanol, reagent on the phosphate ("direct" mode) 4 bases × TEAH⁺ hplc_ternary_2026-10-20/
Bridged ternary reagent chains on C8, residue held only by the reagent ("ionic" mode) residue A × 5 reagents hplc_bridged_2026-10-22/

Bridged ternaries, residue A (kcal/mol; R = residue, T = reagent, S = C8 proxy; ΔE₃ = three-body term):

Reagent R···T T···S R···S ΔE₃
TEAH⁺ (triethylammonium) −19.2 −2.1 0.0 +0.02
DIPEA-H⁺ −20.0 −2.6 −1.0 −0.30
dibutylammonium −23.4 −2.4 0.0 +0.01
hexylammonium −18.3 −3.7 −0.1 +0.01
tributylammonium −24.7 −3.2 0.0 +0.03

For comparison, the silanol-bound TEAH⁺ ternaries give R···S −13.8 to −16.7 and ΔE₃ +1.0 to +1.2 for all four bases.

Pair interactions in the two adsorption modes

Comparison with published IP-RP retention data

Reference data: Donegan, Nguyen & Gilar 2022 (13 alkylamines, oligo-dT on C18; ref. 35) and Maurer et al. 2025 (13 alkylamines, 200–6000 nt RNA on a polymeric DNAPac RP column; ref. 36). For the five reagents modelled here, both rank retention TEA < DIPEA < DBA < HA < TBA.

EDA descriptors vs published retention

What the EDA supports (qualitative):

  1. Length-based separation in the ionic mode. The ion-pair energy per phosphate barely depends on the base (TEAH⁺: A −18.8, C −19.7, G −19.6, U −20.0 kcal/mol). Each phosphate contributes about the same binding, so when retention runs through the reagent the column effectively counts charges. This matches both papers: hydrophobic reagents give length-based separation, and RNA shows on-off elution across 200–6000 nt.
  2. A molecular picture of mixed-mode retention. In the bridged complex the residue never touches the ligand (R···S ≈ 0) and the energies are pairwise additive (ΔE₃ ≈ 0): the oligonucleotide is held at arm's length by the reagent. Only the direct arrangement puts the nucleobase in contact with the surface, and only that can discriminate sequences. This is consistent with Donegan's finding that weak reagents keep sequence, modification and phosphorothioate-diastereomer selectivity, and strong ones suppress it.
  3. Pairing strength does not set the reagent ranking. Dibutyl- and tributylammonium pair 4–6 kcal/mol more strongly than TEAH⁺, but hexylammonium pairs weakest and is among the most retained (rank correlation 0.4). This supports choosing reagents by hydrophobicity (retention, boiling point, logD), as both papers do.
  4. The reagent's N–H is contested. Every isolated reagent···C8 pair optimized into an N–H···silanol hydrogen bond (1.75–2.04 Å). That bond gives most of the −10.6 to −12.7 kcal/mol, while chain-only contact gives just −2 to −4. Maurer's column has no silanols, yet it shows the same reagent order, so the measured order comes from hydrophobic adsorption, not from polar surface sites.

What the current model gets wrong:

  • Reagent order is not reproduced quantitatively. With n = 4–5 reagents, one starting geometry each, and differences of 0.3–2 kcal/mol, no correlation is established (chain-contact T···S in the bridged model: Spearman ρ = 0.8, p = 0.10). C-PCM has no cavitation (hydrophobic) term, but adding the SMD non-electrostatic (CDS) term does not fix this: in water it adds only −0.7 to −1.0 kcal/mol to chain contact (4–6 cal/mol/Ų of buried nonpolar area, vs ≈ 25 from alkane transfer) and changes no rank (reports/hplc_nonpolar_2026-10-26/). On a silanol-free three-chain C8 patch every reagent binds by chain contact (−4.9 to −8.9 kcal/mol, dispersion-like) and TEAH⁺ is weakest. With five placements per reagent (xTB-screened; lowest-energy = most strongly bound in every reagent) the order is TEAH⁺ < DIPEA-H⁺ < HAA < TBuAA < DBAA (ρ = 0.7, p = 0.19; 0.6 with CDS) against the published TEA < DIPEA < DBA < HA < TBA: TBuAA and DBAA both bind by inserting over the chain triangle, but DBAA is still 1.7 kcal/mol stronger and HAA ranks too low (reports/hplc_patch_2026-10-27/, reports/hplc_patch_sampling_2026-10-28/). This order is robust to the solvent model at fixed geometry. Bondi × 1.4 radii bind every pair 0.6 kcal/mol more strongly on average (ρ 0.5–0.7), and an acetonitrile dielectric changes the EDA by < 0.03 kcal/mol (reports/hplc_patch_sensitivity_2026-11-10/).
  • Base order is inverted. Both papers find T > A > G > C, with C/G-rich oligos eluting first. Here the direct residue···C8 contact puts C strongest (−18.9) and A/U weakest (−14.9), because these contacts are polar hydrogen bonds to the silanol.
  • Not modelled: HFIP (Donegan: tertiary amines gain the most apparent hydrophobicity with it), counter-anions, acetonitrile, reagent concentration and micelles (Maurer's 100–200 mM plateau), RNA secondary structure, and reagent mixtures (Maurer's best: 100 mM BA + 50 mM TPA).

Next steps: use a larger/denser brush (≥ 5–7 chains) that linear and tertiary reagents can both insert into, then test whether HA/DBA/TBA > DIPEA/TEA (and TBA > DBA) and T > A > G > C emerge; compare phosphorothioate with phosphate ion pairs; model HFIP pairing with each reagent; add BA, TPA and DEA.

Install

Use Python 3.12 in an isolated environment; do not upgrade system Python/drivers. CPU: python -m pip install -r requirements-cpu-lock.txt -e '.[test]'. The actual host GPU installation and exact pins are in requirements-gpu-lock.txt after verified local execution. Install with python -m pip install -r requirements-gpu-lock.txt -e '.[test]'. CUDA wheel names depend on driver/runtime compatibility. CUDA 12 runtimes can coexist with the host CUDA 13 toolkit. Source scripts/gpu-env.sh before GPU execution; it exposes the environment's NVIDIA wheel libraries without modifying the system loader. The script targets the verified Python 3.12 .env layout. Exact host installation:

/home/zyliu/anaconda3/bin/conda create -p "$PWD/.env" python=3.12 pip -y
.env/bin/python -m pip install -r requirements-gpu-lock.txt -e '.[test]'
source scripts/gpu-env.sh

The optimization foundation passed nine scientific checks with actual GPU execution. Experimental CPU EDA preparation adds independent electrostatic and correlation checks; a validated full GKS-EDA release remains unavailable. See the recorded evidence.

source scripts/gpu-env.sh # GPU wheel library paths
.env/bin/eda-gpu inspect
.env/bin/eda-gpu singlepoint examples/water.xyz --metadata examples/water.json --backend cpu --output runs/water_sp
.env/bin/eda-gpu optimize examples/water_dimer.xyz --metadata examples/water_dimer.json --backend gpu --output runs/dimer_gpu
.env/bin/eda-gpu optimize examples/water.xyz --metadata examples/water_frozen.json --backend cpu --output runs/frozen
.env/bin/eda-gpu validate --gpu --output runs/validation
.env/bin/python -m pytest -q

Output directories must be new, preserving existing work. Failed calculations write result.json with an error and return nonzero. GPU failures never fall back to CPU. validate --gpu requires GPU execution; without it GPU validation is explicitly absent.

JSON metadata requires identity, protonation, integer charge and spin (Nalpha-Nbeta, not multiplicity), and a complete fragment partition with charges/spins. Atom indices are zero based; geomeTRIC constraint files use one based indices. Bonds, stereochemistry, and other provided chemical metadata are preserved in results. XYZ alone cannot establish protonation or stereochemistry. No conformer generation or HPLC retention prediction is implemented.

Results record energies in hartree, gradients in hartree/bohr, angstrom coordinates, SCF/optimization status, method settings, runtime versions/hardware, constraints, initial/final XYZ and evaluation trajectories. Convergence does not establish a minimum.

See methods, validation, GKS-EDA source gate, and the dated report under reports/.

Scientific references, research papers, and websites

If EDA_GPU has benefited your research, please cite this project with the version or commit used, the GKS-EDA publications, and the software, functional, and basis-set papers relevant to your calculation. The bibliography below covers the current implementation, documented benchmarks, and supporting software. Each entry explains its role.

GKS-EDA theory and reference program

  1. Su, P.; Li, H. (2009). “Energy decomposition analysis of covalent bonds and intermolecular interactions.” Journal of Chemical Physics 131, 014102. DOI: 10.1063/1.3159673. Supplies the orbital intermediate energy expressions used by the experimental implementation. University of Nebraska institutional record.
  2. Su, P.; Jiang, Z.; Chen, Z.; Wu, W. (2014). “Energy Decomposition Scheme Based on the Generalized Kohn–Sham Scheme.” Journal of Physical Chemistry A 118, 2531–2542. DOI: 10.1021/jp500405s. Original GKS-EDA method paper and the project's intended decomposition.
  3. Su, P.; Tang, Z.; Wu, W. (2020). “Generalized Kohn-Sham energy decomposition analysis and its applications.” WIREs Computational Molecular Science 10, e1460. DOI: 10.1002/wcms.1460. Review used to map KS orbital intermediates, component definitions, and generalized correlation.
  4. Tang, Z.; Song, Y.; Zhang, S.; Wang, W.; Xu, Y.; Wu, D.; Wu, W.; Su, P. (2021). “XEDA, a fast and multipurpose energy decomposition analysis program.” Journal of Computational Chemistry 42, 2341–2351. DOI: 10.1002/jcc.26765. Reference-program publication; the current comparisons use published manual data.

Official reference resources:

Source-access history and equation provenance are in the source audit and mathematical specification. The original 2014 full text has not been authenticated locally; the supplied review and the Su–Li formulas support the experimental construction. No local XEDA reference execution has been established. Citations and internal energy closure do not establish external component validation. Historical transcription and author-upload links are preserved in those audit documents with their access and reliability limitations.

Published molecular benchmark

  1. Liu, Z.; Trindle, C. O.; Gu, Q.; Wu, W.; Su, P. (2017). “Unravelling hydrogen bonding interactions of tryptamine–water dimer from neutral to cation.” Physical Chemistry Chemical Physics 19, 25260–25269. DOI: 10.1039/C7CP03491A. Source of the neutral/radical-cation M06-2X/aug-cc-pVDZ comparison and Table 3 component values in the attributed benchmark fixture. Publisher article and supporting-information navigation; four-page supporting information. The supplied SI has no complete Cartesian geometries; comparisons using the constructed seed remain exploratory.

Electronic structure, optimization, and numerical software

  1. Sun, Q., et al. (2018). “PySCF: the Python-based simulations of chemistry framework.” WIREs Computational Molecular Science 8, e1340. DOI: 10.1002/wcms.1340. CPU electronic-structure foundation.
  2. Sun, Q., et al. (2020). “Recent developments in the PySCF program package.” Journal of Chemical Physics 153, 024109. DOI: 10.1063/5.0006074. PySCF methods and program development. Also see Sun, Q., et al. (2026), “The Python simulations of chemistry framework: 10 years of an open-source quantum chemistry project,” Journal of Chemical Physics 165, 102502, listed on the official citation page.
  3. Sun, Q. (2015). “Libcint: An efficient general integral library for Gaussian basis functions.” Journal of Computational Chemistry 36, 1664–1671. DOI: 10.1002/jcc.23981. Gaussian-integral library underlying PySCF.
  4. Li, R.; Sun, Q.; Zhang, X.; Chan, G. K.-L. (2024). “Introducing GPU-acceleration into the Python-based Simulations of Chemistry Framework.” arXiv:2407.09700. GPU4PySCF reference listed by its official repository.
  5. Wu, X., et al. (2024). “Enhancing GPU-acceleration in the Python-based Simulations of Chemistry Framework.” arXiv:2404.09452. Additional GPU4PySCF reference listed by its official repository.
  6. Wang, L.-P.; Song, C. (2016). “Geometry optimization made simple with translation and rotation coordinates.” Journal of Chemical Physics 144, 214108. DOI: 10.1063/1.4952956. geomeTRIC optimizer and TRIC coordinates; 2022 erratum.
  7. Lehtola, S.; Steigemann, C.; Oliveira, M. J. T.; Marques, M. A. L. (2018). “Recent developments in Libxc — A comprehensive library of functionals for density functional theory.” SoftwareX 7, 1–5. DOI: 10.1016/j.softx.2017.11.002. Exchange-correlation functional library used through the calculation backends.
  8. Harris, C. R., et al. (2020). “Array programming with NumPy.” Nature 585, 357–362. DOI: 10.1038/s41586-020-2649-2. CPU arrays, density algebra, and result processing.
  9. Virtanen, P., et al. (2020). “SciPy 1.0: fundamental algorithms for scientific computing in Python.” Nature Methods 17, 261–272. DOI: 10.1038/s41592-019-0686-2. Numerical infrastructure used by the scientific software stack.

Official software websites and implementation documentation:

Independent cross-check program

  1. Smith, D. G. A., et al. (2020). “Psi4 1.4: Open-source software for high-throughput quantum chemistry.” Journal of Chemical Physics 152, 184108. DOI: 10.1063/5.0006002. Independent program for the energy/gradient cross-check (version 1.11 used); website and citation guidance.

  2. Bannwarth, C.; Ehlert, S.; Grimme, S. (2019). “GFN2-xTB—An Accurate and Broadly Parametrized Self-Consistent Tight-Binding Quantum Chemical Method with Multipole Electrostatics and Density-Dependent Dispersion Contributions.” Journal of Chemical Theory and Computation 15, 1652–1671. DOI: 10.1021/acs.jctc.8b01176. Seed refinement before DFT optimization in the HPLC model study (xtb).

  3. Ehlert, S.; Stahn, M.; Spicher, S.; Grimme, S. (2021). “Robust and Efficient Implicit Solvation Model for Fast Semiempirical Methods.” Journal of Chemical Theory and Computation 17, 4250–4261. DOI: 10.1021/acs.jctc.1c00471. ALPB implicit solvation used with GFN2-xTB seed refinement.

Density functionals and counterpoise convention

  1. Boys, S. F.; Bernardi, F. (1970). “The calculation of small molecular interactions by the differences of separate total energies. Some procedures with reduced errors.” Molecular Physics 19, 553–566. DOI: 10.1080/00268977000101561. Ghost-basis counterpoise convention for fragment references and BSSE treatment.
  2. Becke, A. D. (1993). “Density-functional thermochemistry. III. The role of exact exchange.” Journal of Chemical Physics 98, 5648–5652. DOI: 10.1063/1.464913. Three-parameter hybrid construction relevant to B3LYP.
  3. Stephens, P. J.; Devlin, F. J.; Chabalowski, C. F.; Frisch, M. J. (1994). “Ab Initio Calculation of Vibrational Absorption and Circular Dichroism Spectra Using Density Functional Force Fields.” Journal of Physical Chemistry 98, 11623–11627. DOI: 10.1021/j100096a001. B3LYP reference returned by the installed Libxc interface.
  4. Becke, A. D. (1988). “Density-functional exchange-energy approximation with correct asymptotic behavior.” Physical Review A 38, 3098–3100. DOI: 10.1103/PhysRevA.38.3098. B88 exchange contribution to B3LYP/B3LYP5.
  5. Lee, C.; Yang, W.; Parr, R. G. (1988). “Development of the Colle-Salvetti correlation-energy formula into a functional of the electron density.” Physical Review B 37, 785–789. DOI: 10.1103/PhysRevB.37.785. LYP correlation contribution to B3LYP/B3LYP5.
  6. Vosko, S. H.; Wilk, L.; Nusair, M. (1980). “Accurate spin-dependent electron liquid correlation energies for local spin density calculations: a critical analysis.” Canadian Journal of Physics 58, 1200–1211. DOI: 10.1139/p80-159. Local correlation and the explicitly recorded B3LYP variant convention.
  7. Perdew, J. P.; Burke, K.; Ernzerhof, M. (1996). “Generalized Gradient Approximation Made Simple.” Physical Review Letters 77, 3865–3868. DOI: 10.1103/PhysRevLett.77.3865. PBE exchange/correlation underlying PBE0 and independent PBE component checks.
  8. Adamo, C.; Barone, V. (1999). “Toward reliable density functional methods without adjustable parameters: The PBE0 model.” Journal of Chemical Physics 110, 6158–6170. DOI: 10.1063/1.478522. PBE0 global hybrid functional.
  9. Ernzerhof, M.; Scuseria, G. E. (1999). “Assessment of the Perdew–Burke–Ernzerhof exchange-correlation functional.” Journal of Chemical Physics 110, 5029–5036. DOI: 10.1063/1.478401. Companion PBE0 reference returned by the installed Libxc interface.
  10. Zhao, Y.; Truhlar, D. G. (2008). “The M06 suite of density functionals for main group thermochemistry, thermochemical kinetics, noncovalent interactions, excited states, and transition elements: two new functionals and systematic testing of four M06-class functionals and 12 other functionals.” Theoretical Chemistry Accounts 120, 215–241. DOI: 10.1007/s00214-007-0310-x. M06-2X functional used for the tryptamine–water benchmark.

Orbital and auxiliary basis sets

  1. Weigend, F.; Ahlrichs, R. (2005). “Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for H to Rn: Design and assessment of accuracy.” Physical Chemistry Chemical Physics 7, 3297–3305. DOI: 10.1039/B508541A. def2-SVP basis used in the optimization-foundation validation.
  2. Weigend, F. (2002). “A fully direct RI-HF algorithm: Implementation, optimised auxiliary basis sets, demonstration of accuracy and efficiency.” Physical Chemistry Chemical Physics 4, 4285–4291. DOI: 10.1039/B204199P. Correlation-consistent J/K-fitting auxiliary bases relevant to the aug-cc-pVDZ-JKFIT density-fitted optimization path.
  3. Dunning, T. H., Jr. (1989). “Gaussian basis sets for use in correlated molecular calculations. I. The atoms boron through neon and hydrogen.” Journal of Chemical Physics 90, 1007–1023. DOI: 10.1063/1.456153. Correlation-consistent cc-pVDZ foundation.
  4. Kendall, R. A.; Dunning, T. H., Jr.; Harrison, R. J. (1992). “Electron affinities of the first-row atoms revisited. Systematic basis sets and wave functions.” Journal of Chemical Physics 96, 6796–6806. DOI: 10.1063/1.462569. Diffuse augmentation underlying aug-cc-pVDZ in the molecular benchmark.
  5. Hehre, W. J.; Ditchfield, R.; Pople, J. A. (1972). “Self—Consistent Molecular Orbital Methods. XII. Further Extensions of Gaussian—Type Basis Sets for Use in Molecular Orbital Studies of Organic Molecules.” Journal of Chemical Physics 56, 2257–2261. DOI: 10.1063/1.1677527. 6-31G basis underlying the XEDA CO₂ example.
  6. Hariharan, P. C.; Pople, J. A. (1973). “The influence of polarization functions on molecular orbital hydrogenation energies.” Theoretica Chimica Acta 28, 213–222. DOI: 10.1007/BF00533485. Polarization functions relevant to the example's 6-31G* basis.
  7. Weigend, F. (2008). “Hartree–Fock exchange fitting basis sets for H to Rn.” Journal of Computational Chemistry 29, 167–175. DOI: 10.1002/jcc.20702. Universal J/K-fitting auxiliary basis family used as def2-universal-jkfit in the optimization-foundation calculations.

Basis definitions and their source references can also be inspected through PySCF basis documentation and the Basis Set Exchange. Spherical/Cartesian representation, auxiliary basis, and installed functional definitions must be reported with each result because these choices affect comparison with published data.

HPLC retention literature (comparison data)

  1. Donegan, M.; Nguyen, J. M.; Gilar, M. (2022). “Effect of ion-pairing reagent hydrophobicity on liquid chromatography and mass spectrometry analysis of oligonucleotides.” Journal of Chromatography A 1666, 462860. DOI: 10.1016/j.chroma.2022.462860. Source of the % acetonitrile at elution and the reagent hydrophobicity order used in the HPLC comparison.
  2. Maurer, J.; Malburet, C.; François-Heude, M.; Guillarme, D. (2025). “Evaluation of ion pairing reversed-phase liquid chromatography for the separation of large RNA molecules.” Journal of Chromatography A 1740, 465574. DOI: 10.1016/j.chroma.2024.465574. Reagent classification and large-RNA retention on a silanol-free polymeric column.

Project terms

This repository is public for development visibility. No project license has been assigned yet. Published literature and manuals are linked; full-text papers are not redistributed in this repository. Software projects retain their own licenses and citation guidance.

Experimental CPU EDA preparation

The supplied 2020 review was read completely. Its unambiguous definitions now support frozen electrostatics and generalized correlation, using direct integrals and ghost-basis counterpoise references at the complex geometry. Exactly two closed-shell, all-electron fragments and B3LYP/PBE0 are supported by this older CPU-only command. Use eda below for the complete experimental CPU/GPU component calculation.

.env/bin/eda-gpu eda-prepare examples/water_dimer.xyz --metadata examples/water_dimer.json --backend cpu --output runs/eda_dimer
.env/bin/eda-gpu eda-prepare examples/co2_dimer_xeda_ii.xyz --metadata examples/co2_dimer_xeda_ii.json --backend cpu --basis '6-31g*' --output runs/eda_co2

A successful preparation reports full_eda_available: false. Exchange, repulsion, polarization, and total closure remain null until their intermediate-state conventions and component references are established. Saved densities support further development. The official XEDA II CO₂ example differs from our partial calculation beyond printed rounding; its settings are not fully specified, so it is an exploratory comparison. See comparison evidence, mathematical specification, and source audit. Papers remain local.

Complete experimental CPU/GPU GKS-EDA

eda now computes electrostatic, exchange, repulsion, polarization (including charge transfer), generalized correlation, and explicit dispersion (zero for the supported non-dispersion-corrected methods). Exactly two all-electron fragments with additive spins, B3LYP/B3LYP5/PBE0, counterpoise references and direct integrals are supported. It uses the explicit Su–Li orbital intermediates on KS orbitals, as specified in the supplied review; no ALMO-EDA substitution is made.

source scripts/gpu-env.sh
.env/bin/eda-gpu eda examples/water_dimer.xyz --metadata examples/water_dimer.json --backend gpu --output runs/gks_dimer_gpu
.env/bin/eda-gpu eda examples/water_dimer.xyz --metadata examples/water_dimer.json --backend cpu --output runs/gks_dimer_cpu
.env/bin/eda-gpu eda examples/co2_dimer_xeda_ii.xyz --metadata examples/co2_dimer_xeda_ii.json --backend gpu --basis '6-31g*' --cart --output runs/gks_co2_gpu
.env/bin/python scripts/validate_gks.py --gpu --repeats 3 --output runs/gks_checks

All terms, intermediate energies, condition checks and reference densities are saved. Gas-phase results state all_components_computed: true, scientifically_validated: true (published-benchmark agreement, ≤0.20 kcal/mol per component) and full_eda_available: false (reserved for a matched-settings XEDA comparison, not performed). Solvated results state scientifically_validated: false. A successful process means the calculation converged and its internal checks passed; it does not imply agreement with XEDA tighter than ~0.2 kcal/mol.

28 local tests passed, including actual GPU EDA, independent integral/PBE0 checks and a many-electron determinant energy check. Repeated CPU/GPU comparisons on water and CO₂ dimers have a maximum individual-component difference of 1.82e-8 hartree. Cartesian functions explain most of the earlier CO₂ reference discrepancy, but residual component differences remain beyond printed rounding.

GPU4PySCF direct J/K has CPU/GPU hybrid internals; this is not an all-GPU workflow. For these tiny cases, measured complete EDA GPU medians were slower than CPU (24.43 versus 3.15 seconds for water dimer; 29.47 versus 9.19 for CO₂). These measurements do not establish larger-system throughput or speedup. See GPU EDA validation and timings, methods, and validation details.

Experimental GKS-EDA(sol)

eda --solvent cpcm|iefpcm [--solvent-eps E] adds the solvent extension of the review (Eqs. 12–13): all reference SCFs use a PCM reaction field, the six gas-phase-form terms are evaluated with the gas-phase functional at the SCRF densities, and ΔG_desol is the change in electrostatic solvation energy. Monomers sit in the supermolecule cavity with only their own nuclei. Two definitions are documented interpretations pending the original derivation (specification). singlepoint and optimize accept the same flags (PCM analytic gradients pass finite-difference checks).

.env/bin/eda-gpu eda examples/water_dimer.xyz --metadata examples/water_dimer.json --backend gpu \
  --xc pbe0 --solvent cpcm --output runs/gks_sol_dimer
.env/bin/python scripts/validate_gks_sol.py --gpu --output runs/gks_sol_validation

Validated internally: vacuum limit (1.6e-9 Eh), dielectric screening at separation, CPU/GPU agreement for C-PCM and IEF-PCM (≤9.6e-9 Eh). Not externally validated: the XEDA II BH₃–NH₃ example differs strongly and depends on unknown cavity radii, to which desolvation is very sensitive. See validation.

Component-level external agreement is established only to the published-table level (≤0.20 kcal/mol on reconstructed geometries, Liu et al. 2017); the project owner accepted this on 2026-10-19 and deferred a matched-settings XEDA run. Tighter agreement than ~0.2 kcal/mol per component is therefore not demonstrated. Gas-phase results report scientifically_validated: true on this basis; full_eda_available stays false. Open-shell, solvent and double-hybrid validation remain deferred; the RNA/HPLC application is exploratory (HPLC application study).

Tryptamine–water paper benchmark (under development)

The benchmark follows Liu et al., PCCP 2017, DOI 10.1039/C7CP03491A: neutral M06-2X and radical-cation U-M06-2X with aug-cc-pVDZ and counterpoise EDA. The cation is created by removing one electron, not by protonation. Restricted and unrestricted CPU/GPU optimization and complete experimental EDA are supported. Open-shell fragment spins must add to the complex spin; broken-symmetry/antiferromagnetic coupling is outside the implemented scope. S² is recorded for unrestricted SCFs.

Published Table 3 data are attributed in references/tryptamine_water_2017.json. The supplied four-page SI contains molecular figures and charges, but no full Cartesian coordinates. examples/tryptamine_water/constructed_bridge.xyz is an explicitly labelled approximate seed constructed from the SI bonding pattern, not the paper's original geometry. Optional seed generation requires rdkit==2026.9.1; ordinary calculations do not require RDKit.

source scripts/gpu-env.sh
.env/bin/python scripts/compare_tryptamine.py \
  --xyz examples/tryptamine_water/constructed_bridge.xyz \
  --metadata examples/tryptamine_water/neutral.json \
  --case A_H2O --opt-density-fitting --output runs/tryptamine_comparison

This runs CPU/GPU density-fitted optimization (aug-cc-pVDZ-JKFIT) sequentially from identical input, then direct counterpoise EDA on the same GPU-optimized geometry. Outputs include coordinates, trajectories, settings, spin diagnostics, aligned geometry RMSD, all EDA terms, backend differences and differences from the paper. Pauli is exchange plus repulsion, matching Table 3. Paper agreement remains exploratory until original coordinates and numerical settings are matched. No ZPE/deformation-corrected D0, minimum/Hessian certification, intra-EDA or individual-H-bond decomposition is claimed. A separate H⁺–H₂O starting conformer is required to study that isomer.

Live calculation progress and research attribution

.env/bin/eda-gpu monitor runs/my-calculation
.env/bin/eda-gpu status runs/my-calculation

monitor follows SCF cycles, optimization evaluations, energies, gradient norms, temperature cooldowns and final EDA components. Ctrl+C exits the viewer while the calculation continues. status prints one snapshot; --json gives machine-readable output. For older running jobs, add --log /path/to/calculation.log. New calculations record atomic progress.json snapshots and an events.jsonl history beside results.

Calculation output, monitoring and JSON results carry this acknowledgement: “If EDA_GPU has benefited your research, please support us by citing this project and the GKS-EDA publications below.” Method attribution includes Peifeng Su et al., original GKS-EDA, 2014, Su, Tang and Wu review, 2020, and Su and Li orbital derivation, 2009. Tryptamine runs additionally cite Liu et al., 2017.

Computing limits

Molecular CLI commands default to 16 physical cores/compute threads, 50% workload pacing, GPU cooldown at 75°C until 65°C and CPU cooldown at 85°C until 75°C. These settings affect this calculation process. --threads, --cpu-core-limit, --compute-duty-cycle, --gpu-temperature-limit and --cpu-temperature-limit are configurable; temperature limit 0 disables that sensor guard. Library Settings retain unpaced defaults for explicitly controlled benchmarks. CPU sensor availability and actual pacing/temperature statistics are reported. Software checks occur at SCF cycles and workload boundaries; long kernels cannot be interrupted by these controls. Pacing is an average workload target, not a fixed instantaneous utilization or watt cap.

GPU power/clocks are never changed automatically. nvidia-smi -pl changes the shared GPU and affects other projects. The developer's original RTX 4080 SUPER power limit was 320 W; restoring it is an explicit administrator action.

For an optional temporary system-wide CPU frequency cap:

sudo /usr/bin/python3 scripts/cpu_frequency_cap.py apply --max-mhz 3000 --minutes 60
sudo /usr/bin/python3 scripts/cpu_frequency_cap.py restore

The helper saves the original limits, applies the ceiling across CPU policies and starts an automatic restore timer. Manual restore can end it early; later manual policy changes are preserved. This hardware setting affects other CPU applications while active and is not installed as a boot service. Neither this helper nor process pacing modifies firmware or Windows configuration.

Completed exploratory neutral tryptamine–water benchmark

GPU M06-2X/aug-cc-pVDZ direct counterpoise GKS-EDA at the GPU density-fitted optimized geometry converged with all six components. Values below are kcal/mol, compared with PCCP Table 3:

Component GPU Paper
Electrostatic -17.279 -17.34
Pauli 19.535 19.69
Polarization -7.048 -7.07
Generalized correlation -5.489 -5.54
Total -10.282 -10.27

This is encouraging exploratory agreement, not a verified reproduction. The maximum component difference is 0.155 kcal/mol. Original Cartesian coordinates are unavailable; the seed was reconstructed. Paper grid settings are unspecified. CPU comparison and cation calculations remain pending. SCF thresholds were 1e-9 Eh and orbital-gradient 1e-6; an earlier 1e-12 Eh attempt did not converge and was excluded. Energy closure residual was 8.45e-11 Eh. See archived results.

About

Under development: experimental CPU/GPU GKS-EDA (six components, counterpoise, RKS/UKS, M06-2X) with DFT gradients and geometry optimization. Energies, gradients and minima cross-checked against Psi4; tryptamine–water totals match the 2017 paper within 0.012 kcal/mol. Matched XEDA component validation pending.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages