In Loving Memory of
Warren Joseph Hehre (1945 - 2026)
Devoted husband, father, mentor, friend.
The world is less clever in his absence.
Machine Learning in Spartan
Neural Networks trained to reproduce quantum-chemical results. Application in specific computational tasks dramatically reduces computation times. Present generation Spartan includes:
DLFF03. The “Corrected MMFF” model. D for Deep L for Learning FF for Force Field (03 = 3rd version). The DLFF03 model was trained on ≈140 thousand ωB97X-V/6-311+G (2df,2p)[6-311G*] dft energies. This network is applicable to multiple standard computational tasks: equilibrium geometry, equilibrium conformer, and is of particular utility in the conformer distribution and flexible molecule NMR Spectrum protocols (MLXD and MLHF).
J. Comp. Chem. 2025, 46, (1), 70016.
Practical Machine Learning Strategies 1.
DLXD2XV, DLXD2MV, DLXD2M2. The “Estimated Density Functional” Energy models, utilized for Boltzmann weighting of conformers in the ML-NMR protocols. model. D for Deep L for Learning XD for ωB97X-D 2(to) XV for ωB97X-V. Similarly, MV, M2 for ωB97M-V or ωB97M(2) functionals. Trained on up to 295,000 high-quality energy calculations (per network) and providing conformational energy differences within 1 kJ/mol of their training models, these three networks are of utility improving energies from the default DFT geometry option: ωB97X-D/6-31G*.
J. Comp. Chem. 2025, 46 (12), 70129.
Practical Machine Learning Strategies 2.
DLFFG1. The “Estimated Density Functional” Geometry model, where G1 reference “geometry version 1”. This neural network was a substantial undertaking including training to more than 6 million energy and force calculations and reproduces bond-distances from ωB97X-D/6-31G* to within 0.002 Å RMS and 0.046 Å max absolute deviation. This dramatically reduces times for “good” structure, a requisite for accurate NMR shifts.
J. Chem. Inf. Model. 2025, 65 (5), 2314 - 2321
Practical Machine Learning Strategies 3.
DLNMR1. The “Estimated DFT NMR Shifts” model. Comprising two separate networks (one for H one for C) and trained against more than 44 million proton and more than 33 million C shifts. Shown to reproduce extremely high quality empirically corrected C NMR shifts within
0.3 ppm, in conjunction with DLFFG1, capable of providing sub 2.0 RMS error C shifts for rigid molecules in seconds.
J. Org. Chem. 2025, 90 (32), 11478-11485.
Practical Machine Learning Strategies 4.
DLHF2XD. This neural network corrects HF energies to ωB97X-D/6-31G* level and serves as a bridge between the calibrating QM HF/3-21G and the DLXD2XV Boltzmann energy step in the MLHF-NMR protocol.
[manuscript in preparation]
