This page highlights methodologies that we developed which create modeling capabilities some of which are unique in the world, and some key results. This is a selection, and it includes only topics where I was the lead
Computing accurate vibrational spectra of molecules, especially of molecules at interfaces, and vibrations in materials, is important for the assignment of species and reaction pathways. This is indispensable specifically in the design of functional materials such as catalysts. For the vast majority of molecule-surface systems as well as liquids and solids that have ever been studied theoretically/computationally, only uncoupled harmonic (normal mode) vibrations are computed, whose accuracy is of the order of 100 cm-1 - hardly enough for reliable assignment. This is an outstanding issue specifically for molecule-surface systems and heterogeneous catalysis where catalytic bond weakening implies enhanced anharmonic effects, but methods to compute them are very imperfect (Phys. Chem. Chem. Phys., 24, 15158-5172 (2022)). Computing spectra with the experimental resolution (~1 cm-1) requires the inclusion of anharmonicity and coupling. This requires solving the vibrational Schrödinger equation with multiple degrees of freedom, for systems where potential energy surfaces might not exist. We have developed the rectangular collocation approach (J. Chem. Theory Comput., 19, 1641–1656 (2023) ) that is able to compute anharmonic spectra directly from a relatively small number of PES samples computable directly ab initio, with small adaptable (J. Chem. Theory Comput., 8, 2053-2061 (2012); Chem. Phys. Lett. 511, 434-439 (2011) ) basis sets and exact kinetic energy operator (J. Chem. Phys., 145, 224110 (2016) ). This facilitates, in particular, calculations of anharmonic spectra of molecule-surface systems important for catalysis (Surf. Sci., 605(5-6), 616-622 (2011); Phys. Chem. Chem. Phys., 15, 10028-10034 (2013); J. Chem. Phys. 139(5), 051101 (2013); MRS Proceedings 1484, 1-6 (2012) ) and other systems where tradictional accurate computational spectroscopy is difficult to apply (J. Phys. Chem A, 119, 9557-9567 (2015) ). We showed that intelligent collocation point placemend, including rule-based placemend and ML-based placement, can reduce the cost or improve the accuracy of the method (Chem. Phys., 509, 139–144 (2018), J. Phys. Chem. A, 123, 10631-10642 (2019) ).
We have developped methods and tools for realiable machine learning in high-dimensional feature spaces and under sparse sampling (Phys. Chem. Chem. Phys., 25, 1546-1555 (2023) ). We use representations of multi-dimensional functions with lower-dimensional functions such as expansions over orders of coupling (J. Chem. Phys. 125, 084109 (2006); J. Phys. Chem. A, 124, 7598−7607 (2020) ; Artif. Intell. Chem., 1, 100013 (2023); ibid. 1, 100008 (2023) ) or redundant coordinates (J. Chem. Phys. 127, 014103 (2007); J. Phys. Chem. A, 127, 7823–7835 (2023) ) to achieve robust ML even with very sparse samping (J. Math. Chem., 61, 7–20 (2023) ) while generating insight about the type and functional forms of the dependence of the target properties on the features (Comput. Phys. Commun., 271, 108220 (2022); Digital Discovery, 3, 1967-1979 (2024); J. Mater. Inform., 5, 38 (2025). We developed a hybrid neural network (NN) - kernel regression method (GPR-NN) that combines the advantages of the high expressive power of an NN with the robustness of kernel regression (J. Phys. Chem. A, 127, 7823–7835 (2023) ), in particular, effectively addressing overfitting. The method allows to use ML to guide the construction of analytic formulas (Mach. Learn. Sci. Technol., 6, 035002 (2025) ). We also proposed codes implementing our methods (Comput. Phys. Commun., 180, 2002-2012 (2009); ibid. 271, 108220 (2022); J. Phys. Chem. A, 127, 7823–7835 (2023) ).
We were the first to show, back in 2006 (J. Phys. Chem. A, 110, 5295 – 5304 (2006) ) that neural networks (NN), far from being a curios but mediocre-accuracy approach that it was mostly believed to be at that time, are able to produce very accurate, spectroscopically accurate PES. We were the first to show that NNs can easily build sum-of-product (SOP) representations that are extremely important for facilitating integrals in quantum dynamics (J. Chem. Phys. 125, 194105 (2006) ). We were the first to compare NN and Gaussian process regression for spectroscopic PES under the same conditions (J. Chem. Phys., 148, 241702 (2018) ). We showed that enforcing of symmetry is not a categoric requirement even for spectroscopically accurate PES (J. Chem. Phys., 159, 211103 (2023) ), justifying the use of generic ML methods and tools for PES construction. We were the first to produce an ab initio-based ML PES for a polyatomic molecule on a catalytic surface explicitly considering all molecular degrees of freedom (Surf. Sci. 604(5-6), 555-561 (2010) ).
We developped machine-learned models of kinetic energy functional (KEF) for orbital-free DFT (OF-DFT) based on NN and kernel regressions (Phys. Chem. Chem. Phys., 21, 378-395 (2019) ; J. Chem. Phys., 153, 074104 (2020); ibid., 159, 234115 (2023); Electr. Struct., 6, 045002 (2024) ) and a code where they can be implemented specifically meant for ML-based functionals (Comput. Phys. Commun., 256, 107365 (2020) ). I proposed the first an analytic KEF guided by ML (arXiv:2502.05411 (2025) ), a development enabled by .the (GPR-NN) method that I developed (J. Phys. Chem. A, 127, 7823–7835 (2023).
We were the first to propose machine-learned local pseudopotentials (LPP) for OF-DFT, inlcuding the use of LPP shape to palliate deficiencies elsewhere in the method (Chem. Phys. Lett., 622, 99-103 (2015); J. Phys. Chem. A, 124, 11111–11124 (2020) ).
We proposed a density functional tight-binding - molecular mechanics (DFTB-MM) hybdrid approach whereby some interatomic interactions are treated at the SCC DFTB level and some at the MM level (J. Chem. Theory Comput., 19, 5189-5198 (2023) ). While traditional QM-MM approaches use QM or MM level of treatment based on spacial domains, DFTB-MM is effectively a type of QM-MM where the level of treatment is by specific interatomic contributions. This allows expanding the area of applicability of DFTB to cases where good parameters are not available (such as many heterostructures).
We showed that it is possible to obtain qualitatively correct bandstructure (such as localized gap state) at the GGA level where GGA+U or hybrid functionals are believed to be required (J. Power Sources, 278, 197-202 (2015) ) and that in general much smaller (and therefore less perturbing to the structure, convergence etc) Hubbard U values can be used, or avoided altogether, with LCAO type bases than with plane waves (MRS Commun., 10, 259-264 (2020) ; Int. J. Quantum Chem., 121, e26439 (2021) ).
We were among the first to explore the use of GGA+U with small Hubbard U corrections to all types of states (i.e. beyond the traditional application to d and f states), for example to achieve correct energy differences between polymorphs, which is at the limit of DFT accuracy, or to achieve correct energetic ordering of electronic states (AIP Adv., 6, 045116 (2016); Chem. Phys. Lett., 659, 270-276 (2016); ChemElectroChem, 7, 3151-3159 (2020) ). This enables cheaper, larger-scale calculations with non-expensive, smaller-basis set GGA+U while ensuring key elements of more accurate calculations.
We were the first to predict that amorphous silicon Si can act as anode in Na-ion batteries (Comput. Mater. Sci., 94, 214-217 (2014) ). The prediction was confirmed experimentally shortly after (Electrochimica Acta, 211, 265-272 (2016) ). We first proposed to use amorphization to increase the voltage or to enable intercalation of metal cations, and provided first truly comparative ab initio studies of the interaction of different cations used in metal-ion batteries (Li, Na, K, Mg, Ca, Al) with different phases, including amorphous of several materials explored as electrode materials for post-Li batteries (Comput. Mater. Sci., 94, 214-217 (2014) ; J. Power Sources, 278, 197-202 (2015); J. Phys. Chem. C 119, 13496–13501 (2015); J. Chem. Phys., 143, 204701 (2015); MRS Adv., 1, 3069–3074 (2016); ibid., 3, 3507-3512; ibid., 4, 837-842 (2019); MRS Commun., 7, 819-825 (2017); Phys. Chem. Chem. Phys., 19, 22538-22545 (2017); ibid., 19, 6076-6081 (2017) J. Phys. D: Appl. Phys., 53, 083001 (2020) ).
We were the first to propose the use of p-doping to increase the voltage or to enable electrochemical activity, in inorganic and organic battery materials (Solid State Ionics, 253, 157-163 (2013) ; J. Power Sources, 274, 65-70 (2015); J. Chem. Phys., 146, 034706 (2017) ; Phys. Chem. Chem. Phys., 19, 13195-13209 (2017); MRS Commun., 7, 523-540 (2017); ibid., 10, 259-264 (2020); Int. J. Quantum Chem., 121, e26439 (2021) ).
We were the first to model organic Li and Na ion battery p-type cathodes in solid state (Phys. Chem. Chem. Phys., 20, 232-237 (2018) ).
We demonstrated conceptually and computationally the ubuquitous nature of oxygen redox, insufficiencly of the commonly used PDOS and formal oxidation state-based analyses to quantify it, and the utility of electron density-based descriptions (J. Phys. Chem. Lett., 8, 1593-1598 (2017); ibid. 8, 3945-3946 (2017); J. Phys.: Conf. Ser. 1136, 012017 (2018) ; J. Phys. Chem. C, 124, 19962−19968 (2020); Molecules, 26, 5541 (2021) ; Chem. Rev., 124, 12661–12737 (2024) ).