21 April 2023 in Physical Chemistry Chemical Physics
We investigated the various factors impacting the performance of Δ-machine learning (Δ-ML) solution phase molecular properties.
30 June 2025 in Digital Discovery
We introduce an activity-cliff explanation supervision training strategy to enhance both predictivity and explainability for graph neural networks in molecular structure and activity relationship modeling.
23 October 2025 in Journal of Chemical Theory and Computation
Time-dependent density functional theory (TDDFT) combined with range-separated hybrid (RSH) functionals and a tuned range-separation parameter γ offers a computationally economical approach for high-throughput excited-state property predictions. The γ-tuning procedure in the gas phase is well established. However, no agreement on the best γ-tuning procedure has been made when considering the solvent effect with implicit solvent models like the polarizable continuum model (PCM). To answer that question, this study created a diverse dataset with 937 molecules with experimental solution-phase UV/vis absorption spectra. Three γ-tuning methods, the gas-phase γ-tuning (GPγT), the partial vertical γ-tuning (PVγT), and the strict vertical γ-tuning (SVγT), were evaluated for the ωPBEh functional over the entire dataset. Additional benchmarks are done for the optimally tuned screened range-separated hybrid combined with the PCM approach (SRSH-PCM) and the solvation-mediated tuning procedure (sol-med-OT). Our findings revealed that the optimal γ-values obtained by the PVγT and the SVγT are significantly smaller than the GPγT. This trend holds consistently across all molecules in our dataset, and we explained the origin of this phenomenon. TDDFT calculations with PVγT- and SVγT-tuned γ-values and default global Fock exchange fraction achieve superior performance compared to those using GPγT-tuned or default γ and slightly outperform SRSH-PCM and sol-med-OT with similar or lesser computational cost. Furthermore, we found that the smaller γ-values from SVγT captured the expected 1/(εR) asymptotic behavior in the solution phase, resulting in accurate prediction of solution-phase CT excitations, consistent with the screened asymptote behavior encoded in SRSH-PCM. These results show that SVγT is the best scheme for high-throughput UV/vis absorption spectrum calculations using the ωPBEh functional from a data-driven perspective.
3 February 2023 in ChemRxiv
Due to the limitation of solvent models, quantum chemistry calculated solution-phase molecular properties often deviates from experimental measurements. Recently, ∆-machine learning (∆-ML) was shown to be a promising approach to correcting errors in the quantum chemistry calculation of solvated molecules. However, this approach's applicability to different molecular properties and its performance in various use cases are still unknown. In this work, we tested the performance of Δ-ML in correcting redox potential and absorption energy calculations using four types of input descriptors and various ML methods. We sought to understand the dependence of ∆-ML performance on the property to predict, the quantum chemistry method, the data set distribution/size, the type of input features, and the feature selection techniques. We found that ∆-ML can effectively correct the errors in redox potentials calculated by density functional theory (DFT) and absorption energies calculated by time-dependent DFT. For both properties, the ∆-ML corrected results showed less sensitivity to the DFT functional choice than the raw results. The optimal input descriptor depends on the property, regardless of the specific ML method used. The solvent-solute descriptor (SS) is the best for redox potential, whereas the combined molecular fingerprint (cFP) is the best for absorption energy. A detailed analysis of the feature space and the physical foundation of different descriptors well explained these observations. Feature selection did not further improve the Δ-ML performance. Finally, we analyzed the limitation of our Δ-ML solvent effects approach in data sets with molecules of varying degrees of electronic structure errors.
30 September 2025 in ChemRxiv
Time-dependent density functional theory (TDDFT) combined with range-separated hybrid (RSH) functionals and a tuned range-separation parameter γ offers a computationally economical approach for high-throughput excited state property predictions. The γ-tuning procedure in the gas phase is well established. However, no agreement on the best γ-tuning procedure has been made when considering the solvent effect with implicit solvent models like the polarizable continuum model (PCM). To answer that question, this study created a diverse dataset with 997 937 molecules with experimental solution-phase UV/vis absorption spectra. Three γ-tuning methods, the gas phase γ-tuning (GPγT), the partial vertical γ-tuning (PVγT), and the strict vertical γ-tuning (SVγT), were evaluated for the ωPBEh functional over the entire dataset. Additional benchmarks are done for the optimally tuned screened RSH combined with the PCM approach (SRSH-PCM), and the solvation-mediated tuning procedure (sol-med-OT). Our findings revealed that the optimal γ-values obtained by the PVγT and the SVγT are significantly smaller than the GPγT. This trend holds consistently across all molecules in our data set, and we explained the origin of this phenomenon. TDDFT calculations with PVγT and SVγT tuned γ-values and default global Fock exchange fraction achieve superior performance compared to those using GPγT-tuned or default γ, and slightly outperform SRSH-PCM and sol-med-OT with similar or lesser computational cost. Furthermore, we found that the smaller γ values from SVγT captured the expected 1/εr asymptotic behavior in the solution phase, resulting in accurate prediction of solution-phase CT excitations, consistent with the screened asymptote behavior encoded in SRSH-PCM. These results show that SVγT is the best among those three schemes for high-throughput UV/vis absorption spectrum calculations using with the ωPBEh functional from a data-driven perspective.
19 December 2024 in ChemRxiv
Computational modeling of the excited states of molecular aggregates faces significant computational challenges and size heterogeneity. Current machine learning (ML) models, typically trained on specific-sized aggregates, struggle with scalability. We found that the exciton model Hamiltonian of large aggregates can be decomposed into dimer pairs, allowing an ML model trained on dimers to reconstruct Hamiltonians for aggregates of any size. We trained an ML model on perylene dimer exciton Hamiltonians using quantum chemistry descriptors to resolve phase correction issues. Our model accurately predicted excitation energies and oscillator strengths in perylene trimers and tetramers, and analyzed the size-dependent optical bandgap in aggregates with up to 50 monomers, revealing quantum dot-like confinement effects. Future work will explore transferability across different monomers to predict optical properties in heterogeneous assemblies.