betaFeedback

WorksPhysics and Astronomy

Semi-Automated Creation of Density Functional Tight Binding Models Through Leveraging Chebyshev Polynomial-based Force Fields

Abstract

Density Functional Tight Binding (DFTB) is an attractive method for accelerated quantum simulations of condensed matter due to its enhanced computational efficiency over standard Density Functional Theory approaches. However, DFTB models can be challenging to determine for individual systems of interest, especially for metallic and interfacial systems where different bonding arrangements can lead to significant changes in electronic states. In this regard, we have created a rapid-screening approach for determining systematically improvable DFTB interaction potentials that can yield transferable models for a variety of conditions. Our method leverages a recent reactive molecular dynamics force field where many-body interactions are represented by linear combinations of Chebyshev polynomials. This allows for the efficient creation of multi-center representations with relative ease, requiring only a small investment in initial DFT calculations. We have focused our workflow on TiH2_2 as a model system and show that a relatively small training set based on unit-cell sized calculations yields a model accurate for both bulk and surface properties. Our approach is easy to implement and can yield accurate DFTB models over a broad range of thermodynamic conditions, where physical and chemical properties can be difficult to interrogate directly and there is historically a significant reliance on theoretical approaches for interpretation and validation of experimental results.

Phone: +1-925-422-3994

Introduction

Atomistic computer simulations are frequently used as an independent route to determining physical and chemical properties for complex systems, including material synthesis,[1] highly strained conditions,[2] and surface chemistry.[3] These calculations can provide predictions of physical properties and reaction mechanisms, which can make experiments more tractable by aiding in their interpretation and helping to narrow the number of different conditions to investigate. Accurate modeling of the breaking and forming of chemical bonds in condensed phases frequently requires the use of quantum theories such as Kohn-Sham Density Functional Theory (DFT). DFT remains one of the most widely used theoretical methods in condensed matter physics, computational chemistry, and materials science for prediction of material properties and chemical reactivity, including phase boundaries and thermal decomposition (e.g., Ref. 4). DFT calculations, though, require immense computational effort per simulation time step and consequently are usually limited to picosecond time scales and nanometer system sizes. In contrast, many processes of interest have properties that can span orders of magnitude larger scales, including large-scale carbon heterocycle synthesis,[5] the rational design of 3D materials,[6] and defect formation and grain boundary interactions in crystalline systems.[7] Hence, there is a great need to explore methods that can harness the accuracy of DFT while yielding substantial improvements in computational speeds.

Density Functional Tight Binding (DFTB) is a semi-empirical quantum theory based on an expansion of the Kohn-Sham DFT total energy that holds promise as an alternate approach for gas-phase studies[8] as well as condensed phase systems.[9, 10] DFTB implements a balance of approximate quantum mechanics and empirical functions that can allow for several orders of magnitude improvement in computational efficiency for a variety of materials and mixtures.[11] Difficulty can arise in creating new DFTB interaction potentials or improving upon existing models due to the large number of hyper-parameters within the approach. In particular, the approximate Hamiltonian in DFTB can be sensitive to the confining potential chosen for the wavefunction and electron density[12] (§ 2.2), and in general there does not exist a predefined recipe for how to choose these parameters nor how to explore that specific phase space. The Hamiltonian parameters in turn are closely coupled to the empirical repulsive energy, which itself has a wide variety of options in terms of functional form and data to be fit. [13] In addition, the repulsive energy is usually taken to be strictly pairwise (two-center), though a number of systems can require many-body terms as well for accurate predictions. [14] The generation of training data can be problematic for high-ZZ or magnetic materials, [15] where sufficient DFT-MD data can be too computationally demanding to compute and being able to leverage small MD training sets would be advantageous. In addition, machine learning approaches tend to be highly data intensive, [16] which can pose difficulties for any method that leverages these techniques. Thus, DFTB method development would be holistically improved through a more automatic method for parameterization, where candidate models could be screened rapidly and efficiently, thereby allowing the user to quickly determine an optimal model for their specific needs.

To this end, we propose a method for streamlining DFTB parameterization by leveraging the recently developed Chebyshev Interaction Model for Efficient Simulation (ChIMES). [17, 18] ChIMES is a many-body reactive force field model for molecular dynamics (MD) simulation based on linear combinations of Chebyshev polynomials. ChIMES MD models have been created for a number of systems under both unreactive and reactive conditions, including molten liquid carbon, [17] water under ambient and high pressure-temperature conditions, [19, 20] high pressure C/O systems, [21, 22] and detonating energetic materials. [23] The main advantage of the ChIMES formalism is that optimal parameters are determined extremely rapidly (e.g., within minutes for this effort) through solving a set of simultaneous linear equations. This in contrast to bond order methods (e.g., (Refs. 24,25)) and neural network force fields (e.g., (Refs. 8,26)) that generally require non-linear optimization (e.g., Levenberg-Marquardt), making determination of optimal parameter sets significantly more labor intensive and uncertain due to possible local minima in the search space.

In this work, we use the metal TiH2_2 for our model determination. TiH2_2 has a number of possible applications areas, including in hydrogen storage alloys, [27] and as a superconducting material, [28] a thermal energy storage material, [29], an intermediary to produce fine Ti metal powders, [30, 31] and as a blending agent to create porous aluminum foams. [32] It exhibits face centered cubic (fcc) symmetry in its ground state, with the (111) surface generally exhibiting the lowest surface energy (Figure 1). Small scale TiH2_2 calculations are relatively tractable with DFT, facilitating thorough validation of any developed DFTB parameter set. However, DFT calculations are too computationally cumbersome to model the grain boundaries and density of crystalline defects that likely play a significant role in material performance in these application areas. These studies would thus greatly benefit from a high throughput quantum approach that can determine larger-scale effects that are challenging to determine from DFT alone. In addition, study of TiH2_2 serves as a starting point for determining (de)hydriding reactions in similar metal hydrides such as ZrH2_2, which have a number of uses such as in solid propellants [33] and nuclear fuel cladding tubes. [34] The modeling framework developed here is general and can easily be applied to magnetic alloys such as Fe/Ni mixtures, which have use in magnetic refrigeration systems, magnetic recording devices, and other applications. [35]

Pictures of TiH2 bulk and surfaces

Figure 1. Pictures of TiH2_2 bulk and surfaces. The left panel shows the bulk fcc lattice. The middle panel shows the (111) crystalline surface the Top (marked with an ‘O’) and Hollow (‘X’) adsorption sites indicated. The right panel shows the (011) crystalline surface with the Top (‘O’), Bridge-1 (‘X1_1’) and Bridge-2 (‘X2_2’) sites indicated.

We begin with a brief discussion of our DFT calculations and the DFTB method, followed by details of the ChIMES formalism. We then experiment with different options for the DFTB quantum potential energy and charge transfer terms, and indicate optimal choices for our system. Next, we proceed through our workflow for DFTB model selection, beginning with the search for optimal values for the DFTB confining potentials, followed by the same for the ChIMES polynomials orders, and finally some discussion of optimal least-squares fitting practices. In all cases, we deliberately use a relatively small DFT training set as a somewhat stringent test for our results, and to assess the possibility of fitting to sparse data.

Methods

DFT Calculations

All DFT calculations were performed with the Vienna ab initio Simulation Package (VASP)[36, 37, 38] using projector-augmented wave function (PAW) pseudopotentials[39, 40] and the Perdew-Burke-Ernzerhof exchange correlation functional (PBE).[41] We found our results to be converged with a planewave cutoff of 400 eV and an energy convergence criteria of 10−610^{-6} eV, both of which were used for the results reported here. Fourth order Methfessel-Paxton smearing[42] was used with a value of 0.13 eV for all geometry and cell lattice optimizations in order to ensure energy convergence without dependence on the electronic smearing temperature. The Mermin functional[43] with the same electronic temperature was used for all MD calculations in order to avoid spurious forces due to possible negative occupation numbers from the Methfessel-Paxton approach. Brillouin Zone sampling for all TiH2_2 unit cell calculations was performed with a 10×10×1010 \times10 \times10 k-point mesh, whereas we used a mesh of 5×5×55 \times5 \times5 for 32 formula unit (96 atom) bulk calculations. We used system sizes of 168 atoms/7 layers for the (001) surface, 144 atoms/6 layers for the (011) surface, and 192 atoms/8 layers for the (111) surface, each with a vacuum of 20 Å and a k-point mesh of 5×5×15 \times5 \times1 in the direction of the surface.

Our DFT training set consisted of molecular dynamics simulations of unit cell configurations (12 atoms total), run for 5 ps at 400 K with simulation cells initially optimized to pressures of −8, 0, 10, and 100 GPa. All MD calculations were run in the constant temperature and volume (NVTNVT) ensemble with Nosé-Hoover thermostat chains[44, 45, 46] and a timestep of 0.2 fs. The slightly elevated temperature and wide pressure range including negative pressure were chosen in order to yield a broad sampling of the underlying potential energy surface. Atomic forces and the diagonal of the stress tensor were then sampled from MD configurations at fixed time intervals of ∼160\sim160 fs in order ensure configurations were as statistically uncorrelated as possible. This yielded up to 30 MD snapshots for each pressure.

In addition, in order to sample hyper- and hypo-coordinated configurations in the system, we included MD data for a unit cell with a single hydrogen interstitial or single vacancy site, each run for 5 ps. The starting configuration for these defect simulations were taken from the ground-state stoichiometric unit cell, with a single H atom removed or added in an octahedral interstitial site. This yielded a total of 153 unit cell-sized configurations for our training set.

DFTB theory and calculations

The formalism for DFTB with self-consistent charges (SCC) has been discussed in detail elsewhere. [11, 47, 48, 49, 50] Briefly, the method assumes neutral, spherically symmetric, atom-centered charge densities n0n_0 and that the true charge density of the system is within a small perturbation, i.e., n(r)=n0(r)+δnn(\mathbf{r}) = n_0(\mathbf{r}) + \delta n. In its most typical form, a second-order Taylor expansion of the Kohn-Sham total energy in δn\delta n is employed to yield the following energy expression:

E[δn]≈∑afa⟨ψa∣T^+V^[n0]∣ψa⟩+12∬(δ2Exc[n0]δnδn′+1∣r−r′∣)δnδn′−12∫n0′n0∣r−r′∣+Exc[n0]−∫Vxc[n0]n0+EII,(1)\begin{aligned} E[\delta n] \approx{}& \sum_a f_a \left\langle\psi_a \left| \hat{T} + \hat{V}[n_0] \right| \psi_a \right\rangle\\ &+ \frac{1}{2} \iint\left( \frac{\delta^2 E_{\mathrm{xc}}[n_0]}{\delta n \delta n'} + \frac{1}{\lvert\mathbf{r} - \mathbf{r}' \rvert} \right) \delta n \delta n' \\ &- \frac{1}{2} \int\frac{n_0' n_0}{\lvert\mathbf{r} - \mathbf{r}' \rvert} + E_{\mathrm{xc}}[n_0] - \int V_{\mathrm{xc}}[n_0] n_0 + E_{\mathrm{II}}, \tag*{(1)} \end{aligned}

where we have employed a shorthand of ∫=∫dr\int= \int d\mathbf{r}, ∫′=∫dr′\int' = \int d\mathbf{r}', n0(r)=n0n_0(\mathbf{r}) = n_0, etc. The first line in (1) is the band structure energy:

EBS[n0]=∑afa⟨ψa∣H[n0]∣ψa⟩,(2)E_{\mathrm{BS}}[n_0] = \sum_a f_a \left\langle\psi_a \left| H[n_0] \right| \psi_a \right\rangle, \tag*{(2)}

where the Hamiltonian HH does not contain any charge transfer terms. EBSE_{\mathrm{BS}} is thus computed via sum over occupied electronic states determined from diagonalization of the approximate DFTB Hamiltonian. The Hamiltonian and overlap matrices used in EBSE_{\mathrm{BS}} are determined from pre-tabulated Slater-Koster (SK) tables derived from calculations with a minimal, non-orthogonal basis set.

Prior to SK tabulation, both the electronic wave functions and electron density are subjected to separate confining or compression potentials. Most commonly, the confining potential has the general form of:

Vconf(r)=(rRχ)β.(3)V_{\mathrm{conf}}(r) = \left(\frac{r}{R_{\chi}}\right)^{\beta}. \tag*{(3)}

The exponent β\beta is usually set to a value of 2, and Rχ=RψR_{\chi}=R_{\psi} or RnR_{n}, where RψR_{\psi} corresponds to wavefunction compression and RnR_{n} to density compression. Our past experimentation with larger values of β\beta yielded negligible difference in the final outcome, though this could be investigated systematically in future efforts. The compression potentials force the wavefunction/electron density to zero at relatively large distances from the nuclei, which has been shown to improve transferability of the SK tabulations. [47] Choice of a short wavefunction compression improves transferability by removing the overly diffuse parts of the orbitals. However, this generally comes at the expense of the accuracy of the electron density at longer distances, which can affect the description of non-covalent interactions. [51] Hence, the standard convention is to compute the wavefunction and density compression separately, with Rn>RψR_{n} > R_{\psi}.

The potential energy in the DFTB Hamiltonian is generally expressed in one of two forms. In its original formulation, [52] the Hamiltonian is taken to be a superposition of potentials centered on each atom, which for the purposes of this study we label as the “DFTB2-pot” model:

H^μντr1AB=⟨ϕμA∣T^+V^(nA)+V^(nB)∣ϕνB⟩(4)\hat{H}^{AB}_{\mu\nu\tau r_1} = \left\langle\phi_{\mu}^{A} \left| \hat{T} + \hat{V}(n_{A}) + \hat{V}(n_{B}) \right| \phi_{\nu}^{B} \right\rangle \tag*{(4)}

Here, AA and BB correspond to the atomic index, and μ\mu and ν\nu to an atomic orbital of specific angular momentum. This is akin to an N-body expansion of the quantum mechanical potential energy that is truncated after the one-center or one-body terms, only.

A more common DFTB Hamiltonian involves a superposition of electronic densities instead, which we label as the “DFTB2-dens” model:

H^μνT2AB=⟨ϕμA|T^+V^(nA+nB)|ϕνB⟩(5)\hat{H}^{AB}_{\mu\nu T2} = \left\langle\phi^A_\mu\middle| \hat{T} + \hat{V}(n_A+n_B) \middle| \phi^B_\nu\right\rangle \tag*{(5)}

In this case, the electron density centered on atoms AA and BB is summed before evaluating the potential energy. This amounts to a two-body expansion, where the quantum mechanical potential energy includes both one and two-center terms.

The second line from Equation (1) is the energy from second order charge fluctuations:

ECoul2=12∬(δ2Exc[n0]δnδn′+1∣r−r′∣)δnδn′(6)E_{\mathrm{Coul}}^{2} = \frac{1}{2} \iint\left( \frac{\delta^{2} E_{\mathrm{xc}}[n_{0}]}{\delta n \delta n'} + \frac{1}{\lvert\mathbf{r} - \mathbf{r}' \rvert} \right) \delta n \delta n' \tag*{(6)}

ECoul2E_{\mathrm{Coul}}^{2} is generally referred to as the Coulombic energy, and is evaluated self-consistently using the Hubbard UU parameter. [47, 48] Hubbard UU values for each atom type are precomputed with the uncompressed basis set as the difference between the ionization energy and electron affinity.

In addition to these models, there exists a more recent DFTB energy expression, where third-order density fluctuation terms are included in the Kohn-Sham energy expansion. [49, 53] This yields an additional charge-transfer term to Equation (1), namely,

ECoul3=16∭(δ3Exc[n0]δnδn′δn′′)×δnδn′δn′′(7)E_{\mathrm{Coul}}^{3} = \frac{1}{6} \iiint\left( \frac{\delta^{3} E_{\mathrm{xc}}[n_{0}]}{\delta n \delta n' \delta n''} \right) \times\delta n \delta n' \delta n'' \tag*{(7)}

Evaluation of Equation (7) requires the derivative of the Hubbard UU with respect to charge. [54] This derivative can be determined numerically and for each angular momentum shell of each atom, separately. For our study, we include ECoul3E_{\mathrm{Coul}}^{3} only in conjunction with the density superposition potential (Equation (5)), and label this interaction model as “DFTB3-dens”.

Finally, the last line in Equation (1) is called the repulsive energy:

ERep=−12∫n0′n0∣r−r′∣+Exc[n0]−∫Vxc[n0]n0+EII.(8)E_{\mathrm{Rep}}=-\frac{1}{2}\int\frac{n_0'n_0}{\lvert\mathbf{r}-\mathbf{r}'\rvert}+E_{\mathrm{xc}}[n_0]-\int V_{\mathrm{xc}}[n_0]n_0+E_{\mathrm{II}}. \tag*{(8)}

The nomenclature derives from the ion-ion repulsion term, EIIE_{\mathrm{II}}, though it also contains Hartree double-counting and exchange-correlation terms. In practice, ERepE_{\mathrm{Rep}} is expressed as a short-ranged empirical function that covers first coordination shell or bonded interactions, only. Its parameters are usually fit to reproduce DFT or experimental data, and it can be either pair-wise [13, 55] or contain multi-center interactions. [56, 57]

All DFTB calculations discussed within this work were performed with the DFTB+ code, [11, 50] using self-consistent charges (SCC) [47] and charge convergence criteria of 2.72×10−52.72 \times10^{-5} eV (10−610^{-6} au). Inclusion of an external van der Waals correction [58, 59] is beyond the scope of our present study. We have performed “shell-resolved” SCC calculations, where separate Hubbard UU parameters were determined for each orbital angular momentum shell. Hubbard UU derivatives were computed numerically for DFTB3-dens calculations, where we determined values of −0.0328-0.0328, −0.0840-0.0840, and −0.05334-0.05334 for the ss, pp, and dd-orbital interactions for Ti and a value of −0.1857-0.1857 for the ss-orbital for H. We have also applied an exponential charge damping factor for Ti-H interactions of 3.5, similar to previous efforts. [54] The properties investigated here were largely insensitive to choice of either Hubbard UU derivative or damping factor values. The same electron thermal smearing and k-point mesh were used as described for our DFT calculations for each system.

The ChIMES training set was determined by computing DFTB forces (FF) and diagonal stress tensor components (σ\sigma) for each configuration with the chosen set of Hamiltonian parameters (i.e., {Rψ}\{R_{\psi}\}, {Rn}\{R_{n}\}, DFTB2-pot, DFTB2-dens, or DFTB3-dens interaction) with zero values for those components from ERepE_{\mathrm{Rep}}. These “repulsive energy free” results were then subtracted from the DFT values for those quantities, i.e.,

FRepαiT∗=FDFTαiT−FQM,DFTBαiTσRepααT∗=σDFTααT−σQM,DFTBααT(9)\begin{aligned} F_{\mathrm{Rep}_{\alpha i}}^{\mathrm{T*}} &= F_{\mathrm{DFT}_{\alpha i}}^{\mathrm{T}}-F_{\mathrm{QM,DFTB}_{\alpha i}}^{\mathrm{T}} \\ \sigma_{\mathrm{Rep}_{\alpha\alpha}}^{\mathrm{T*}} &= \sigma_{\mathrm{DFT}_{\alpha\alpha}}^{\mathrm{T}}-\sigma_{\mathrm{QM,DFTB}_{\alpha\alpha}}^{\mathrm{T}} \tag*{(9)} \end{aligned}

Here, τ\tau corresponds to a specific MD configuration, α\alpha to the cartesian direction, and ii is the atomic index. The ‘*’ is used to denote that the quantities being computed are part of the training set, and ‘QM,DFTB’ refers to the quantum components of the DFTB calculation, i.e., only forces and stresses from EBSE_{\mathrm{BS}} and ECoulE_{\mathrm{Coul}}. Inclusion of configurational total energies generally resulted in minimal impact on quality of the ERepE_{\mathrm{Rep}} fit and were thus excluded from our training data, similar to previous efforts.[56, 57] This results in the following objective function:

Fobj=1Nd(∑τ=1M∑i=1N∑α=13[FChIMESαiτ−FRepαiτ∗]2+∑τ=1M∑α=13[σChIMESαατ−σRepαατ∗]2)(10)F_{\mathrm{obj}} = \sqrt{\frac{1}{N_d}\left(\sum_{\tau=1}^{M}\sum_{i=1}^{N}\sum_{\alpha=1}^{3}\left[F^{\tau}_{\mathrm{ChIMES}_{\alpha_i}}-F^{\tau*}_{\mathrm{Rep}_{\alpha_i}}\right]^2+\sum_{\tau=1}^{M}\sum_{\alpha=1}^{3}\left[\sigma^{\tau}_{\mathrm{ChIMES}_{\alpha\alpha}}-\sigma^{\tau*}_{\mathrm{Rep}_{\alpha\alpha}}\right]^2\right)} \tag*{(10)}

where MM is the total number of configurations in the training set, and NdN_d is the total number of data entries (3MN3MN force components plus 3M3M stress tensor components).

ChIMES models and determination of parameters

The design philosophy behind ChIMES involves mapping quantum mechanical energies onto linear combinations of many-body Chebyshev polynomials of the first kind. Chebyshev polynomials of the first kind have a number of desirable properties for creation of interatomic potential energy surfaces, including: (i) they are orthogonal (with respect to a weighting function) and can be generated recursively, allowing for basis set completeness and user defined complexity, (ii) higher order polynomials tend to have decreasing expansion coefficient values (due to their monic form), and (iii) they are “nearly optimal” (the error in an expansion will closely resemble a minimax polynomial). In addition, derivatives of Chebyshev polynomials of the first kind are are related to Chebyshev polynomials of the second kind, which themselves are orthogonal and can be generated recursively. This allows for easy and reliable determination of forces and stress tensor components for atomistic calculations.

Briefly, the ChIMES total energy corresponds to an nn-body expansion:

EnB=∑i1na1Ei1+∑i1>i2na2Ei1i2+∑i1>i2>i3na3Ei1i2i3+⋯+∑i1>i2…inB−1>inBnanBEi1i2…inB,(11)E_{n_{\mathrm{B}}}=\sum_{i_1}^{n_{\mathrm{a}}}{}^{1}E_{i_1}+\sum_{i_1>i_2}^{n_{\mathrm{a}}}{}^{2}E_{i_1i_2}+\sum_{i_1>i_2>i_3}^{n_{\mathrm{a}}}{}^{3}E_{i_1i_2i_3}+\cdots+\sum_{i_1>i_2\ldots i_{n_{\mathrm{B}}-1}>i_{n_{\mathrm{B}}}}^{n_{\mathrm{a}}}{}^{n_{\mathrm{B}}}E_{i_1i_2\ldots i_{n_{\mathrm{B}}}}, \tag*{(11)}

where EnBE_{n_{\mathrm{B}}} is the total ChIMES system energy, nBn_{\mathrm{B}} is the maximum bodiedness, nEi1i2…in{}^{n}E_{i_1i_2\ldots i_n} is the nn-body ChIMES energy for a given set of atoms with indices i={i1,i2,…,in}i=\{i_1,i_2,\ldots,i_n\}, and nan_{\mathrm{a}} is the total number of atoms in the system. The one-body energies, 1Ei1{}^{1}E_{i_1}, correspond to the atomic energy constants for each element type.

The two-body (pairwise) energies are expressed as linear combinations of Chebyshev polynomials of the first kind:

2Ei1i2=fp(ri1i2)+fcei1ei2(ri1i2)∑m=1O2Cmei1ei2Tm(si1i2ei1ei2)(12){}^{2}E_{i_1i_2}=f_{\mathrm{p}}(r_{i_1i_2})+f_{\mathrm{c}}^{e_{i_1}e_{i_2}}(r_{i_1i_2})\sum_{m=1}^{O_2}C_m^{e_{i_1}e_{i_2}}T_m(s_{i_1i_2}^{e_{i_1}e_{i_2}}) \tag*{(12)}

In this case, Tm(si1i2ei1ei2)T_m(s_{i_1i_2}^{e_{i_1}e_{i_2}}) represents a Chebyshev polynomial of order mm, and si1i2ei1ei2s_{i_1i_2}^{e_{i_1}e_{i_2}} is the pair distance transformed to occur over the interval [−1,1][-1,1] using a Morse-like function[60, 61] (See Ref. 17 for details). Here, si1i2ei1ei2∝exp⁡(−ri1i2/λe1e2)s_{i_1i_2}^{e_{i_1}e_{i_2}}\propto\exp(-r_{i_1i_2}/\lambda_{e_1e_2}) and λe1e2\lambda_{e_1e_2} is an element-pair distance scaling constant, usually taken to be the peak position of the first coordination shell. Cmei1ei2C_m^{e_{i_1}e_{i_2}} is the corresponding permutationally invariant coefficient for the interaction between atom types ei1e_{i_1} and ei2e_{i_2}, taken from the set of all possible element types, {e}\{e\}. The term fcei1ei2(ri1i2)f_{\mathrm{c}}^{e_{i_1}e_{i_2}}(r_{i_1i_2}) is a Tersoff cutoff function[24] which is set to zero beyond a maximum distance defined for a given {e1,e2}\{e_1,e_2\} pair set. In order to prevent sampling of ri1i2r_{i_1i_2} distances below what is sampled in our DFT training set, we introduce use of a smooth penalty function fp(ri1i2)f_{\mathrm{p}}(r_{i_1i_2}). We refer the reader to previous work for additional details.[20]

We can now create a greater than two-body orthogonal basis set by taking products of the (n2)\binom{n}{2} unique constituent pairwise polynomials of the higher order terms. In other words, a three-body term has (32)=3\binom{3}{2}=3 pairs, which yields the following expression for the ChIMES three-body energy:

3Ei1i2i3=fcei1ei2(ri1i2)fcei1ei3(ri1i3)fcei2ei3(ri2i3)∑m=0O3∑p=0O3∑q=0O3′Cmpqei1ei2ei3Tm(si1i2ei1ei2)Tp(si1i3ei1ei3)Tq(si2i3ei2ei3).(13){}^{3}E_{i_1i_2i_3}=f_{\mathrm{c}}^{e_{i_1}e_{i_2}}(r_{i_1i_2})f_{\mathrm{c}}^{e_{i_1}e_{i_3}}(r_{i_1i_3})f_{\mathrm{c}}^{e_{i_2}e_{i_3}}(r_{i_2i_3})\sum_{m=0}^{\mathcal{O}_3}\sum_{p=0}^{\mathcal{O}_3}\sum_{q=0}^{\mathcal{O}_3}{}'C_{mpq}^{e_{i_1}e_{i_2}e_{i_3}}T_m(s_{i_1i_2}^{e_{i_1}e_{i_2}})T_p(s_{i_1i_3}^{e_{i_1}e_{i_3}})T_q(s_{i_2i_3}^{e_{i_2}e_{i_3}}). \tag*{(13)}

We take a triple sum for the i1i2i_1i_2, i1i3i_1i_3, and i2i3i_2i_3 polynomials over the hypercube up to O3\mathcal{O}_3, and include a single permutationally invariant coefficient for each set of powers and atom types, Cmpqei1ei2ei3C_{mpq}^{e_{i_1}e_{i_2}e_{i_3}}. We use the primed sum to denote that only terms for which two or more of the m,p,qm,p,q polynomial powers are greater than zero are included in order to guarantee that three distinct atom-centers are evaluated. The expression for 3Ei1i2i3{}^{3}E_{i_1i_2i_3} also contains the fcf_{\mathrm{c}} smoothly varying cutoff functions for each constituent pair distance. Penalty functions are not included in this case and instead are handled entirely by the two-body interaction.

Similarly, the four-body energy can be written as a product of the (42)=6\binom{4}{2}=6 unique pairwise interactions:

4Ei1i2i3i4=fcei1ei2(ri1i2)fcei1ei3(ri1i3)fcei1ei4(ri1i4)fcei2ei3(ri2i3)fcei2ei4(ri2i4)fcei3ei4(ri3i4)×∑m=0O4B∑p=0O4B∑q=0O4B∑r=0O4B∑s=0O4B∑t=0O4B′Cmpqrstei1ei2ei3ei4Tm(si1i2ei1ei2)Tp(si1i3ei1ei3)Tq(si1i4ei1ei4)Tr(si2i3ei2ei3)Ts(si2i4ei2ei4)Tt(si3i4ei3ei4).(14)\begin{aligned} {}^{4}E_{i_1i_2i_3i_4}={}&f_{\mathrm{c}}^{e_{i_1}e_{i_2}}(r_{i_1i_2})f_{\mathrm{c}}^{e_{i_1}e_{i_3}}(r_{i_1i_3})f_{\mathrm{c}}^{e_{i_1}e_{i_4}}(r_{i_1i_4})f_{\mathrm{c}}^{e_{i_2}e_{i_3}}(r_{i_2i_3})f_{\mathrm{c}}^{e_{i_2}e_{i_4}}(r_{i_2i_4})f_{\mathrm{c}}^{e_{i_3}e_{i_4}}(r_{i_3i_4})\\ &\times\sum_{m=0}^{\mathcal{O}_{4\mathrm{B}}}\sum_{p=0}^{\mathcal{O}_{4\mathrm{B}}}\sum_{q=0}^{\mathcal{O}_{4\mathrm{B}}}\sum_{r=0}^{\mathcal{O}_{4\mathrm{B}}}\sum_{s=0}^{\mathcal{O}_{4\mathrm{B}}}\sum_{t=0}^{\mathcal{O}_{4\mathrm{B}}}{}'C_{mpqrst}^{e_{i_1}e_{i_2}e_{i_3}e_{i_4}}T_m(s_{i_1i_2}^{e_{i_1}e_{i_2}})T_p(s_{i_1i_3}^{e_{i_1}e_{i_3}})T_q(s_{i_1i_4}^{e_{i_1}e_{i_4}})T_r(s_{i_2i_3}^{e_{i_2}e_{i_3}})T_s(s_{i_2i_4}^{e_{i_2}e_{i_4}})T_t(s_{i_3i_4}^{e_{i_3}e_{i_4}}). \tag*{(14)} \end{aligned}

Again, the Cmpqrstei1ei2ei3ei4C_{mpqrst}^{e_{i_1}e_{i_2}e_{i_3}e_{i_4}} correspond to permutationally invariant coefficients of linear combination. In this case, the primed sum indicates that three or more of the mm, pp, qq, rr, ss, tt polynomial orders are greater than zero. Thus far we have limited our studies to include up to four-bodied interactions.[23] However, the practice of generating higher-bodied terms through multiplication of the constituent pairwise polynomials can be generalized to include an (n2)\binom{n}{2}-multiple sum over the nn-body hypercube,[18] which is the subject of future work.

Optimal ChIMES parameters (the coefficients of linear combination) can then readily be determined through the overdetermined matrix equation AC=BrepAC=B_{\mathrm{rep}}. The matrix AA corresponds to the values of the requisite polynomials for a given training configuration, or in other words the derivatives with respect to the fitting coefficients. The column vectors CC and Brep\mathbf{B}_{\mathrm{rep}} correspond to the linear ChIMES coefficients and the numerical values for the training set, respectively. This linear least-squares optimization problem can be solved for with any number of algorithms (§ 3.5).

Results and Discussion

Tests of different forms of the DFTB interaction energy

In order to test each of our DFTB interaction models, we have initially set Ti compression radii to RψTi=5.0R_{\psi}^{Ti}=5.0 au and RnTi=14.0R_{n}^{Ti}=14.0 au. These are similar to established values for titanium, and such as the matsci-0-2, [62] trans3d-0-1, [63] and tiorg-0-1 [64] parameter sets. For the entirety of our study, hydrogen compression radii were set to the same values as those for the miomod-hh-0-1 parameter set, namely, RψH=2.5R_{\psi}^{H}=2.5 au and RnH=3.0R_{n}^{H}=3.0 au. We then generated the SK tables for all interactions. For this section, we have computed single-point properties based on the DFT optimized unit cell, e.g., without determining ERepE_{\mathrm{Rep}} and performing a geometry or lattice optimization. In doing so, we can probe the quality of each interaction model through direct one-to-one comparison to results for DFT computed eigenstates without having to be concerned with the quality of the repulsive energy parameterization.

Analysis of the total electronic density of states (DOS) reveals that all three interaction models yield similar curve shapes relative to DFT in terms of peak energy positions and their relative heights (Figure 2). We have centered each DOS curve around its Fermi energy (EFermi=0E_{\mathrm{Fermi}}=0) in order to facilitate comparison. The DFTB2-pot interaction yields the closest comparison to DFT in terms of the overall peak positions and heights, including eigenstates up to ∼3\sim3 eV. However, the agreement between states above this is significantly poorer, where DFTB2-pot completely omits the doublet at ∼5\sim5 eV present in DFT. In contrast, DFTB2-dens and DFTB3-dens show less accurate agreement with DFT for eigenstates below 3 eV, but improved agreement above this value where both models yield the 5 eV doublet.

Figure 2. Total electronic density of states for bulk face centered cubic TiH₂. The dashed corresponds to the result from the DFTB2-pot Hamiltonian, the red line to DFTB2-dens, the blue line to DFTB3-dens, and the shaded area to DFT. Each curve is centered around its Fermi energy.

We have also determined the Mulliken atomic charges for Ti and H for the bulk system (Table 1), with comparison to Bader charges determined from VASP. While determining a direct correspondence between Mulliken and Bader charges is not possible, we observe reasonable agreement between DFTB2-dens and DFTB3-dens with the results from DFT. In contrast, charges computed from DFTB2-pot are several orders of magnitude smaller and of the opposite sign to the other methods. We observe relatively small differences between DFTB2-dens and DFTB3-dens, either for the electronic DOS or for the atomic charges, with DFTB2-dens arguably yielding slightly improved agreement with the DOS from DFT. The ECoul3E_{\mathrm{Coul}}^{3} interaction was developed in part to improve descriptions of polarizability in hydrogen-bonded systems, [65] and consequently its effect could prove more substantial for highly polarizable or charged species. Regardless, for the remainder of this work we will use the DFTB2-dens interaction model.

MethodqTiq_{\mathrm{Ti}}qHq_{\mathrm{H}}
DFT+1.90+1.90−0.95-0.95
DFTB2-pot−6.9e−3-6.9e-3+3.4e−3+3.4e-3
DFTB2-dens+0.32+0.32−0.16-0.16
DFTB3-dens+0.430+0.430−0.215-0.215

Table 1. Comparison of computed atomic charges for bulk TiH2\mathrm{TiH}_2. Bader charges are computed from DFT whereas Mulliken charges are computed from our DFTB model Hamiltonians.

Description of the ERepE_{\mathrm{Rep}} optimization workflow

Given our choice of interaction potential of superposition of densities and second-order ECoul2E_{\mathrm{Coul}}^{2}, we can now create a workflow for determination of optimal DFTB hyper-parameters (Figure 3). This includes evaluation of RψR_{\psi} and RnR_{n} values, two-body or greater polynomial order for the ChIMES ERepE_{\mathrm{Rep}}, and choice of linear optimization and regularization algorithms for calculation of the ChIMES coefficients. Our workflow is fast and efficient, with the most significant amount of time and CPU resources spent on computing the DFT training set. Each loop of our workflow has been automated, with the user being able to make decisions for how to proceed with hyper-parameter adjustments based on results for validation data.

Flowchart showing the creation of DFTB $E_{\mathrm{rep}}$ models through ChIMES force field parameterization. Compute DFT-MD data; Select $\hat{H}$, $R_{\psi}$, $R_n$; Create SKF files; Compute DFTB training set: $\vec{F}_{DFT}-\vec{F}_{DFTB}$ (no $E_{\mathrm{rep}}$), $\sigma_{\alpha\alpha,DFT}-\sigma_{\alpha\alpha,DFTB}$ (no $E_{\mathrm{rep}}$); Choose ChIMES 2B, 3B, 4B orders, cutoff radii and determine $E_{\mathrm{rep}}$; Validation set: Bulk: lattice const., VDOS, 1H and 2H vacancies; (001), (011), and (111) surface energies; $E_{\mathrm{ads}}$ on (011) and (111) surfaces (5 total); (011) and (111) surface and subsurface H vacancy energies (8 total); Desired accuracy achieved? Yes; Complete; No

Figure 3. Flowchart for creation of DFTB ErepE_{\mathrm{rep}} models through ChIMES force field parameterization.

Our initial DFT-MD simulations were run concurrently for 24 hours on four nodes each of a Linux cluster, with thirty-six 2.1 GHz Intel Xeon E5-2695 v4 cores and 128 gigabytes memory per node. This represented a usage of over 20,000 CPU hours. Generation of SK tables for chosen values of RψR_{\psi} and RnR_{n} run were on a single CPU and required approximately one hour of wall-clock time for each set. Again, hydrogen compression radii were held at miomod-hh-0-1 values. For this effort, we choose to use a ‘global’ compression radius for Ti, where all orbitals use the same confinement radius. This is consistent with previous Ti related parameterizations available for download, [62, 63, 64] all which use the same confinement radius for s, p, and d orbitals. There is some possibility that separate confinement for the vacant 4p orbital could yield improved accuracy, [55] which is the subject of future work. Calculation of the ERepE_{\mathrm{Rep}} training set was run using twelve OpenMP threads, requiring ∼30\sim30 minutes.

Solving the linear least-squares problem for a single ChIMES parameterization required less than one minute on a single CPU in all cases, allowing for extremely rapid screening of different sets of parameters. All minimum and cutoff radii for the ChIMES ERepE_{\mathrm{Rep}} were set to include the first coordination shell sampled in our training set, only: 2.5≤rTiTi≤3.52.5 \le r_{\mathrm{TiTi}} \le3.5 Å and 1.5≤rHTi≤2.51.5 \le r_{\mathrm{HTi}} \le2.5 Å. We use values of λTiTi=3.0\lambda_{\mathrm{TiTi}} = 3.0 Å and λHTi=2.0\lambda_{\mathrm{HTi}} = 2.0 Å for the Morse-like coordinate transforms. H-H repulsive interaction were not sampled in our training set and were thus takes from the miomod-hh-0-1 parameter set.

ChIMES parameterization was followed by calculation of validation data, which included results explicitly not present in our training data. The total validation set included the bulk lattice constant and hydrogen vacancy energy, surface energies of the (001), (011) and (111) crystal facets, hydrogen adsorption energies on the (011) and (111) surfaces, and (011) and (111) surface and sub-surface hydrogen vacancy energies. The (001) surface constructed for our calculations contained a contained a net dipole field due to an imbalance of H ions on the top vs. bottom surface for the stoichiometric system, and as such we did not include its hydrogen surface adsorption and vacancy energies in our validation. Calculation of the entire validation set was performed using twelve OpenMP threads and required approximately one hour to complete. Thus, each design loop in our workflow that included generation of SK tables and ERepE_{\mathrm{Rep}} training data took about two and half hours of wall clock time overall.

Sweep of RψR_{\psi} and RnR_{n} values

For this effort, we have computed SK tables using confining radii values ranging from 3.2≤RψTi≤5.03.2 \le R_{\psi}^{Ti} \le5.0 au and 6.0≤RnTi≤17.06.0 \le R_{n}^{Ti} \le17.0 au. Hydrogen compression radii were held at the previously mentioned values. We find that this range of radii spans those normally sampled in DFTB models for low-ZZ elements as well as transition metals. Smaller values of RψTiR_{\psi}^{Ti} and RnTiR_{n}^{Ti} resulted in some distortions to the electronic DOS without yielding additional accuracy for our validation tests. For this section of our study, we used a constant ChIMES 2-body polynomial order of 12 and a 3-body order of 8, similar to previous work. [56, 57] All optimization discussed in here were performed with the LASSO/LARS algorithm with regularization of 10−310^{-3} (see § 3.5 for details). In doing so, we are able to rapidly create a reasonable ERepE_{\mathrm{Rep}} for our models, allowing us to assess the effects of different values of RψTiR_{\psi}^{Ti} and RnTiR_{n}^{Ti} independently.

We find that that there was small variation between results the root mean-square (RMS) errors for the training set, as well as for validation of the bulk lattice constant, vacancy energies, and (011) hydrogen surface adsorption energies. In addition, all models yielded the DFT-predicted surface energy ordering of E111<E011<E001E_{111} < E_{011} < E_{001}. Hence, we focus our validation discussion on the E111E_{111} surface energy, the (E001/E111)(E_{001}/E_{111}) ratio, and the relative energies of adsorption on the (111) Top and Hollow surface sites, (E111Top−E111Holl)\left(E_{111}^{\mathrm{Top}} - E_{111}^{\mathrm{Holl}}\right).

Our analysis reveals an approximate linear correlation between {RψTi,RnTi}\{R_{\psi}^{Ti}, R_{n}^{Ti}\} in terms of the accuracy of the E111E_{111} energy (Figure 4, top panel). The most accurate results for the E111E_{111} surface energy appear to occur when the radii are approximately consistent with each other (either low or high). Small values of RψTiR_{\psi}^{Ti} combined with long values of RnTiR_{n}^{Ti} tend to over-predict E111E_{111}, where {RψTi=3.2,RnTi=17.0}\{R_{\psi}^{Ti}=3.2, R_{n}^{Ti}=17.0\} predicts a surface energy that is ∼20%\sim20\% too large. In contrast, larger values of RψTiR_{\psi}^{Ti} combined with smaller values of RnTiR_{n}^{Ti} tend to under-predict, where {RψTi=5,RnTi=6.0}\{R_{\psi}^{Ti}=5, R_{n}^{Ti}=6.0\} yields a surface energy that is ∼11%\sim11\% too low.

Figure 4. Results for sweep of values of RψTiR_{\psi}^{Ti} and RnTiR_{n}^{Ti}. The top panel corresponds to the fractional deviation of the surface energy, (E111DFTB−E111DFT)/E111DFT\left(E_{111}^{DFTB}-E_{111}^{DFT}\right)/E_{111}^{DFT}, and the middle panel to the deviation of (E001/E111)(E_{001}/E_{111}) relative to DFT. The bottom panel corresponds to the fractional deviation of (E111Top−E111Holl)\left(E_{111}^{Top}-E_{111}^{Holl}\right), where we have shifted the results so that positive values indicate the correct ordering of E111Top>E111HollE_{111}^{Top}>E_{111}^{Holl}. A perfect matching with DFT in this case would yield a value of one.

In contrast, all of models created here under-predict (E001/E111)(E_{001}/E_{111}) relative to our DFT calculations (Figure 4, middle panel). We observe ratios of 1.44 or so for small RψTiR_{\psi}^{Ti} that monotonically decrease to values of ∼1.35\sim1.35 as RψTiR_{\psi}^{Ti} is increased, whereas the DFT predicted value is 1.70. Our results indicate a much smaller dependence on choice of RnTiR_{n}^{Ti}, where altering its value creates little change in the ratio for a given RψTiR_{\psi}^{Ti}. We note that there can be strong dependence of the surface energies on choice of DFT functional (e.g., Ref. 66), although the relative energetic ordering tends to be consistent.

Finally, our models yield a variety of results for (E111Top−E111Holl)(E_{111}^{\mathrm{Top}} - E_{111}^{\mathrm{Holl}}) (Figure 4, bottom panel), where DFT yields top site adsorption energies that are 0.68 eV higher than that for the hollow site. We found that predicting this energy difference was a particularly difficult challenge for our modeling effort. This is in part likely because Top site adsorption yields a hyper-coordinated Ti surface atom with H chemisorption, whereas Hollow site adsorption yields a more energetically favorable physisorption interaction with the H ion at larger distances from the surrounding surface Ti atoms. For this set of ChIMES parameters, we find that large RψTiR_{\psi}^{Ti} and RnTiR_{n}^{Ti} tend to reverse the energetic ordering of adsorption on these sites, with a top site energy that is 0.1 eV lower than that for the hollow site. In contrast, the results for RnTi=6.0R_{n}^{Ti} = 6.0 generally yielded the best agreement with DFT. These results show monotonic decreases with increasing RψTiR_{\psi}^{Ti}, with a value of 0.20 eV for RψTi=3.2R_{\psi}^{Ti} = 3.2 and a value of 0.13 eV for RψTi=5.0R_{\psi}^{Ti} = 5.0.

Based on these results, we choose to proceed with our SK tables generated from {RψTi=3.6,RnTi=6.0}\{R_{\psi}^{Ti} = 3.6, R_{n}^{Ti} = 6.0\}. We find this combination yields a nearly identical E111E_{111} to DFT with a value of 0.080 eV/Å2^{2}, and provides a relatively high (E001/E111)(E_{001}/E_{111}) ratio of 1.41 for this set of ChIMES parameters. In addition, this model provides a reasonable prediction for the top and hollow site adsorption energies, with an energy difference of 0.17 eV. This model set was chosen over {RψTi=3.2,RnTi=6.0}\{R_{\psi}^{Ti} = 3.2, R_{n}^{Ti} = 6.0\} due to the similarity in accuracy between these sets and in order to avoid any unforeseen distortions to the electronic states due to small confining radii.

Sweep of ChIMES polynomial orders

Given our choice of {RψTi=3.6,RnTi=6.0}\{R_{\psi}^{Ti} = 3.6, R_{n}^{Ti} = 6.0\}, we can now experiment with ChIMES polynomial orders in order to refine our model further. In this case, we have looped over our workflow by sampling the 2-body order a range of 8≤O2B≤168 \leq\mathcal{O}_{2\mathrm{B}} \leq16 and the 3-body order over a range of 0≤O3B≤160 \leq\mathcal{O}_{3\mathrm{B}} \leq16. Once again, we present results from LASSO/LARS optimization with regularization of 10−310^{-3}. Each design loop in this case required only ∼1\sim1 hour (since SK table and ERepE_{\mathrm{Rep}} training data generation were complete), almost all of which was spent on validation calculations.

Our results indicate weak correlation between choice of 2-body and 3-body polynomial order (Figure 5). Inspection of results from a single 3-body order (e.g., O3B=4\mathcal{O}_{3\mathrm{B}} = 4) indicates only small changes in any of the validation results presented here with varying values for the 2-body order. We find that 2-body only repulsive energies (with O3B=0\mathcal{O}_{3\mathrm{B}} = 0) yield some improved accuracy in terms of (E111Top−E111Holl)\left(E_{111}^{\mathrm{Top}} - E_{111}^{\mathrm{Holl}}\right), where we observe values of 0.25 eV for {O2B=8,O3B=0}\{\mathcal{O}_{2\mathrm{B}} = 8, \mathcal{O}_{3\mathrm{B}} = 0\} and 0.28 eV for {O2B=16,O3B=0}\{\mathcal{O}_{2\mathrm{B}} = 16, \mathcal{O}_{3\mathrm{B}} = 0\}. However, these 2-body only repulsive energies yield E111E_{111} surface energies that are consistently over 30% too high and an (E001/E111)\left(E_{001}/E_{111}\right) ratio that is over 20% too small (e.g., a deviation of 0.020 eV/A˚2\text{\AA}^{2} and a ratio of 1.35 for {O2B=8,O3B=0}\{\mathcal{O}_{2\mathrm{B}} = 8, \mathcal{O}_{3\mathrm{B}} = 0\}). In contrast, we find that increasing the 3-body orders to non-zero values yields substantially improved E111E_{111} prediction and a somewhat improved (E001/E111)\left(E_{001}/E_{111}\right) ratio (a deviation of <0.001<0.001 eV/A˚2\text{\AA}^{2} and ratio of 1.43 for {O2B=8,O3B=4}\{\mathcal{O}_{2\mathrm{B}} = 8, \mathcal{O}_{3\mathrm{B}} = 4\}). However, large 3-body orders result in somewhat worse values for (E111Top−E111Holl)\left(E_{111}^{\mathrm{Top}} - E_{111}^{\mathrm{Holl}}\right), with values of 0.16 eV for all O3B=16\mathcal{O}_{3\mathrm{B}} = 16 parameter sets.

Figure 5. Results for sweep of 2-body and 3-body ChIMES polynomial orders, with the same definitions as Figure 4.

Lastly, in order to examine the possibility of including higher-body interactions in ERepE_{\mathrm{Rep}}, we have determined a ChIMES parameterization which includes a 4-body order of four, i.e., {O2B=8,O3B=4,O4B=4}\{\mathcal{O}_{2\mathrm{B}} = 8, \mathcal{O}_{3\mathrm{B}} = 4, \mathcal{O}_{4\mathrm{B}} = 4\}. We find that this ChIMES model yields virtually identical results to those from {O2B=8,O3B=4}\{\mathcal{O}_{2\mathrm{B}} = 8, \mathcal{O}_{3\mathrm{B}} = 4\}. For example, both E111TopE_{111}^{\mathrm{Top}} and E111HollE_{111}^{\mathrm{Holl}} agree between both models within 0.01 eV. This is in part due to the fact that the vast majority of the 4-body ChIMES parameters are small and consequently are set to a value of zero by the LASSO/LARS algorithm. This implies that the majority of interactions encompassed in the repulsive energy (for this system, at least) are three-center, only, which can provide guidance for future ERepE_{\mathrm{Rep}} determinations.

As a result, we opt to proceed with a set of {O2B=8,O3B=4}\{\mathcal{O}_{2\mathrm{B}} = 8, \mathcal{O}_{3\mathrm{B}} = 4\} for this part of our effort. This set yields a value of E111=0.080E_{111} = 0.080, which is nearly identical to the DFT result. In addition, we observe values of (E001/E111)=1.43(E_{001}/E_{111}) = 1.43 and (E111Top−E111Hol)=0.19(E_{111}^{\mathrm{Top}} - E_{111}^{\mathrm{Hol}}) = 0.19 eV. These values are similar to those from {O2B=12,O3B=8}\{\mathcal{O}_{2\mathrm{B}} = 12, \mathcal{O}_{3\mathrm{B}} = 8\}, but yield slightly improved agreement to DFT results overall (1.70 and 0.68 eV). In addition, the lower polynomial orders help decrease the risk of overfitting that might not be observed with the current validation set.

Test of linear least-squares solving methods

Given our choices of {RψTi,RηTi}\{R_{\psi}^{T_i}, R_{\eta}^{T_i}\} and {O2B,O3B}\{\mathcal{O}_{2\mathrm{B}}, \mathcal{O}_{3\mathrm{B}}\}, we now explore several different options for the solving the linear least-squares and regularization to obtain the ChIMES coefficients. For these tests, we have used the Singular Value Decomposition (SVD) and Least-Angle Regression (LARS) with LASSO regularization (abbreviated as LASSO/LARS). We now offer a brief discussion of each method and leave details to the pertinent references.

SVD[67] solves for optimal fitting coefficients directly by performing an eigendecomposition of the generally rectangular A\mathbf{A} matrix and computing its pseudo-inverse. Regularization can be performed by setting singular values (eigenvalues of the square matrix in the SVD decomposition) with an absolute value below a given threshold to zero. In our work, we take this parameter to be Dmax⁡ϵD_{\max}\epsilon, where Dmax⁡D_{\max} is the maximum singular value of A\mathbf{A} and ϵ\epsilon is a factor below a value of one.

LARS is a type of forward step-wise or iterative regression. [68, 69] In this case, all model coefficients are initialized to zero and the covariate (i.e., polynomial values) most correlated to the error residual is determined (i.e., those having the most significant impact on the fit). The corresponding ChIMES parameter is modified incrementally to minimize the error residual until a second covariate yields an equal correlation. At this point, it is included in the active parameter set and both coefficients are modified simultaneously. The process continues until all coefficients are included in the solution, at which point a result equivalent to ordinary least squares fitting is obtained. In practice, LARS steps using a subset of all possible parameters are chosen.

LASSO[70] is an L1-norm regularization method whereby regularization is based on the sum of the absolute values of the fitting coefficients, which has the effect of shrinking a subset of parameters to zero. In this case, the objective function FobjF_{\mathrm{obj}} (Equation (10)) is minimized with the following additional constraint:

FobjLASSO=Fobj+2α∑i=1ni∣ci∣.(15)F_{\mathrm{obj}}^{\mathrm{LASSO}} = F_{\mathrm{obj}} + 2\alpha\sum_{i=1}^{n_i} \lvert c_i \rvert. \tag*{(15)}

Here, nin_i is the total number of unique fitting parameters, cic_i. The parameter α\alpha regularizes the magnitude of the fitting coefficients, which reduces possible overfitting. The LASSO method can be implemented as a variant of LARS where parameters are either added or removed at each solution stage. We find the LASSO variant of LARS to be numerically stable for ill-conditioned A\mathbf{A} matrices, which are often found in force matching.

Our exploration spanned SVD ε\varepsilon values between 10−5≤ε≤10−310^{-5} \leq\varepsilon\leq10^{-3} and LASSO/LARS α\alpha values between 10−5≤α≤10−210^{-5} \leq\alpha\leq10^{-2} (Figure 6). Our sweep of LASSO and SVD regularization values and subsequent validation testing help ensure that our DFTB/ChIMES models are not overfit for the conditions of interest here. SVD optimization with ε=10−2\varepsilon= 10^{-2} resulted in unstable models and are subsequently not included in our discussion. Results for E111E_{111} show little dependence on choice of algorithm or regularization parameter. Both optimization algorithms yield very similar values for the E111E_{111} surface energy, with regularization of 10−310^{-3} showing the best agreement with DFT and LASSO/LARS showing near perfect accuracy. In contrast, the other properties discussed here show some variation between algorithms for regularization of 10−310^{-3}. SVD yields a somewhat lower value than LASSO/LARS for the (E001/E111)(E_{001}/E_{111}) ratio (1.418 vs. 1.431) and a slightly higher value for (E111Top−E111Holl)(E_{111}^{\mathrm{Top}} - E_{111}^{\mathrm{Holl}}) (0.21 vs. 0.19 eV). The agreement overall between algorithms is improved for smaller regularization parameters. LASSO/LARS optimization with α=10−2\alpha= 10^{-2} can appear to be an appealing choice because this parameterization maximizes (E111Top−E111Holl)\left(E^{\mathrm{Top}}_{111} - E^{\mathrm{Holl}}_{111}\right). However, this model yields relatively large errors in the TiH2_2 lattice constant, (4.311 Å, compared to a value of 4.440 Å from α=10−3\alpha= 10^{-3} and 4.417 Å from DFT) and E111E_{111} surface energy (0.092 eV/Å2^2, compared to values of 0.080 eV/Å2^2 from α=10−3\alpha= 10^{-3} and DFT). Hence, we choose to proceed with LASSO/LARS optimization with α=10−3\alpha= 10^{-3} as the best choice for our TiH2_2 model, though SVD optimization with ϵ=10−3\epsilon= 10^{-3} is nearly equivalent.

Figure 6. Results for sweep of different linear least-squares optimization approaches, with the same definitions as Figure 4. The purple bar corresponds to LASSO/LARS, and the green to SVD. Validation set calculations with an SVD regularization of 10−210^{-2} were unstable and hence are not included.

Final set of results

Our final set of hyper-parameter values includes {RψTi=3.6\{R^{\mathrm{Ti}}_{\psi} = 3.6 au, RnTi=5.0R^{\mathrm{Ti}}_{n} = 5.0 au}\} and {O2B=8,O3B=4}\{\mathcal{O}_{2\mathrm{B}} = 8, \mathcal{O}_{3\mathrm{B}} = 4\}, optimized with LASSO/LARS and regularization of α=10−3\alpha= 10^{-3}. This model yields RMS errors of 1.76 eV/Å for hydrogen forces, 1.35 eV/Å for titanium forces, and 0.35 GPa for the stress tensor diagonal. We now discuss results for our full set of validation data. For the remainder of our discussion, we refer to our optimal model as “DFTB/ChIMES”.

Results for bulk properties (Table 2) indicate that DFTB/ChIMES yields a lattice constant with errors of only ∼\sim0.4% and 1.0% from DFT and experiment,[71] respectively. However, our model yields a hydrogen bulk vacancy energy (EvacE_{\mathrm{vac}}) that is ∼\sim0.5 eV too small. We found that underestimation of the vacancy energy at about this magnitude was typical for all ChIMES parameterizations created in this work. Systematic errors like these could be rectified through a number of methods, such as use of an expanded basis set,[14] improved descriptions of charge transfer,[72] or inclusion of multi-center terms in the DFTB Hamiltonian,[73] which is the subject of future work.

We find our model yields accurate surface energies for all three low-index facets investigated in this study (Table 3). In particular, the E011E_{011} and E111E_{111} values are nearly identical to those from DFT. The E001E_{001} value from DFTB/ChIMES is around 17% lower than than that for our DFT calculations (0.114 vs. 0.136 eV/Å2^2). However, this could be due in part to the internal electric field on the (001) surface configuration studied here. DFTB generally can underestimate surface electrostatic interactions due to its determination of atom-centered point charges only in Coulombic interactions.[74]

Calculation of surface and sub-surface hydrogen point vacancies for the (111) and (011) facets shows that DFTB/ChIMES is able to yield the same general trend relative to DFT (Table 4). For the (111) surface, DFTB/ChIMES indicates that removal of a first layer surface hydrogen is 0.52 eV higher than that for the second layer (first sub-surface layer), compared to a value of 0.55 eV from DFT. DFTB/ChIMES essentially recovers the bulk vacancy energy in the second layer, similar to DFT where the second layer vacancy energy is only 0.12 eV greater than that of the bulk. For the (011) surface, DFTB/ChIMES predicts that the first layer EvacE_{\mathrm{vac}} is 0.18 eV higher than that of the bulk, followed by a 0.57 eV drop in value in the second layer relative to the first. This is in accordance with our DFT results, which indicates an increase of 0.34 eV for the first layer relative to the bulk followed by a relative decrease of 0.51 eV in the second layer. Both DFTB/ChIMES and DFT indicate that the second layer EvacE_{\mathrm{vac}} is lower than that of the bulk (0.40 eV for DFTB/ChIMES and 0.17 eV for DFT), and that the bulk EvacE_{\mathrm{vac}} is recovered by the third surface layer. A value of 0.5 eV could be added as a correction factor to vacancy calculations on these systems.

SurfaceLayerDFTB/ChIMESDFT
1111st2.983.67
1112nd2.463.12
1113rd2.462.97
1114th2.493.02
0111st2.653.34
0112nd2.072.83
0113rd2.542.99
0114th2.523.00

Table 4. TiH2\mathrm{TiH}_2 surface vacancy energies (in eV). The “layer” label corresponds to the surface layer from which the H atom is removed.

Our DFTB/ChIMES results show similarly strong agreement with hydrogen surface adsorption energies (Figure 1 and Table 5). We observe the correct energetic ordering of adsorption on the (111) Top and Hollow sites, though the Hollow site energy is 0.35 eV smaller than that from DFT. We see similar agreement with DFT for the (011) surface. Here, DFTB/ChIMES show close agreement for Top site adsorption with a difference of only 0.05 eV from DFT. Our model yields Bridge-1 and Bridge-2 adsorption energies that differ from DFT by 0.29 eV and 0.21 eV, respectively, and incorrectly predicts that the Top site is the lowest energetically of the three. However, these values are similar in energy for all surface sites and we have overall favorable agreement.

SurfaceSiteDFTB/ChIMESDFT
111Top-1.888-1.760
111Hollow-2.081-2.440
011Top-2.383-2.332
011Bridge-1-2.154-2.442
011Bridge-2-2.132-2.342

Table 5. Surface hydrogen adsorption energies on TiH2\mathrm{TiH}_2 surface sites (in eV).

In order to further test the transferability of DFTB/CHIMES, we have performed calculations on both α\alpha-Ti (hcp lattice) and a gas-phase TiH4\mathrm{TiH}_{4} molecule, neither of which were part of our training set (Table 6 and Table 7). We find that our DFTB/ChIMES model yields slightly improved accuracy compared to our DFT results for α\alpha-Ti with in errors in the lattice constants of only 6.1% in the aa lattice constant and 0.4% in the cc lattice constant relative to experiment.[75] In addition, DFTB/ChIMES predicts the correct tetrahedral structure for TiH4_4 (H-Ti-H bond angle of 109.47°) with a Ti-H bond length of 1.76 Å, compared to our DFT computed value of 1.71 Å. Finally, DFTB/ChIMES predicts an energy of reaction for TiH4→_4 \rightarrow TiH3_3 + H of 2.42 eV, compared to the DFT computed value of 2.88 eV. We note the difference of 0.46 eV nearly exactly corresponds to the systematic error in our hydrogen vacancy calculations.

Methodaacc
DFTB/ChIMES3.134.67
DFT2.824.55
Expt. [75]2.954.69

Table 6. α\alpha-Ti lattice constants in Å.

MethodH-Ti-H bond angle (degrees)Ti-H bond length (Å)H dissociation energy (eV)
DFTB/ChIMES109.471.762.42
DFT109.471.712.88

Table 7. TiH4_4 gas-phase molecular data.

Finally, we have computed the vibrational density of states (VDOS) from MD calculations of bulk ground-state TiH2_2 at 400 K (Figure 7). We have made comparison to DFTB/ChIMES results using the unit cell and k-point mesh of 10×10×1010 \times10 \times10, as well as comparison to simulation of a 3×3×23 \times3 \times2 supercell (216 atoms total) using a 2×2×22 \times2 \times2 k-point mesh. Our results indicate strong agreement between DFTB/ChIMES and DFT for the manifold of both Ti-Ti and H-Ti vibrational modes. In particular, we see a close match between relative peak heights and positions for the DFT and DFTB/ChIMES unit cell simulations at frequencies less than 400 cm−1^{-1} and greater than 1000 cm−1^{-1}, including the small shoulder at ∼\sim1150 cm−1^{-1}. We observe some small shifting of peaks when comparing the DFTB/ChIMES unit cell and replicated supercell results, though these changes are generally <10< 10 cm−1^{-1} and are thus energetically quite small (i.e., <10−3< 10^{-3} eV). Overall, we find our DFTB/ChIMES model to yield robust agreement with DFT, allowing for MD calculations with larger system sizes that could incorporate more realistic defect concentrations and/or grain boundaries.

Figure 7. Vibrational density of states calculated from an MD simulation of bulk TiH₂ at 400 K. Unit cell simulations using DFT (black) and our DFTB/ChIMES model (red) are compared to results from the periodically replicated supercell (blue). Note that the oscillations at low frequency are artifacts of the Fourier transform for the small timescale and system size of that simulation.

Conclusions

In this work, we have created a workflow for DFTB model creation that allows for nearly exhaustive search for optimal Hamiltonian parameters and repulsive energies. Our efforts leverage the ChIMES reactive force-field, based on linear combinations of Chebyshev polynomials for the determination of the DFTB repulsive energy. Optimal ChIMES parameters can be determined extremely rapidly to create many-body interaction potentials, which facilitates testing of any number of DFTB related parameters. Each design loop is run in a semi-automated fashion, allowing for a systematic exploration of the manifold of options for the confining potential radii, polynomial order and bodied-ness for the ChIMES potential, and type of least-squares algorithm or regularization to solve for the ChIMES parameters. In doing so, we are able to accelerate the down selection of DFTB models for a given application, removing the need for ad hoc parameter choices.

Our results indicate DFTB/ChIMES models can be accurately determined based on relatively small training data (unit cell MD calculations in this work), even for physically complex systems such as those containing surface chemistry. Further refinement of our TiH2_2 model could involve inclusion of training data from additional phases and thermodynamic state points. A systematic expansion of our training set could be performed by first generating an approximate model, and then performing DFT calculations on selected snapshots and refitting ChIMES repulsive parameters until convergence is reached, [76] similar to our previous force field fitting efforts. [17] Regardless, our current effort yields accurate results for bulk and surface TiH2_2 properties, and our model shows strong transferability to bulk α\alpha-Ti and TiH4_4. The small training set could yield significant advantages for computationally challenging systems such as magnetic materials and their interfaces, where DFT data is limited and difficult to generate. Overall, our DFTB/ChIMES approach can have particular impact on myriad of research areas, such as interpretation of imaging and spectroscopy experiments on bulk and interfacial systems, where there is traditionally a strong coupling with atomistic simulation approaches.

Acknowledgments

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. Computations were performed at LLNL using the borax, boraxo, quartz, rztrona, and rztopaz massively parallel computers. Projects 18-SI-001 with Jason Jeffries as PI and 20-SI-004 with Brandon Wood as PI were funded by the Laboratory Directed Research and Development Program at LLNL. The ChIMES force field code is available for download at https://github.com/rklindsey/chimes_calculator (last accessed on April 13th, 2021). The DFG grant RTG 2247 is acknowledged. ChIMES potential parameters are available upon request.

References

MethodLattice constant (Å)H vacancy energy (eV)
DFTB/ChIMES4.4002.47
DFT4.4173.00
Expt. [71]4.45–

Table 2. TiH2\mathrm{TiH}_2 bulk properties.

SurfaceDFTB/ChIMESDFT
1110.0800.080
0110.1050.101
0010.1140.136

Table 3. TiH2\mathrm{TiH}_2 surface energies (in eV/Ų).

TOC Graphic showing the workflow from Compute DFT through Select DFTB parameters, Compute DFTB training set, Choose ChIMES parameters, Compute validation set, Desired accuracy?, Complete, Yes, and No, with Bulk, (111) surface, (011) surface, and The ChIMES Calculator

Figure 8 (PDF p. 45). TOC Graphic

References

  1. [1]Chandrakumar, K. R. S.; Page, A. J.; Irle, S.; Morokuma, K. Carbon Coating Precedes SWCNT Nucleation on Silicon Nanoparticles: Insights from QM/MD Simulations. J. Phys. Chem. C 2013, 117, 4238–4244.DOI
  2. [2]Steele, B. A.; Goldman, N.; Kuo, I.-F. W.; Kroonblawd, M. P. Mechanochemical synthesis of glycine oligomers in a virtual rotational diamond anvil cell. Chem. Sci. 2020, 11, 7760–7771.DOI
  3. [3]Li, S.; Rangarajan, S.; Scaranto, J.; Mavrikakis, M. On the structure sensitivity of and CO coverage effects on formic acid decomposition on Pd surfaces. Surface Science 2021, 709, 121846.DOI
  4. [5]Kroonblawd, M. P.; Lindsey, R. K.; Goldman, N. Synthesis of Nitrogen-Containing Polycyclic Aromatic Hydrocarbons in Impacting Glycine Solutions. Chemical Science 2019, 10, 6091.DOI
  5. [6]Sours, T.; Patel, A.; Norskov, J.; Siahrostami, S.; Kulkarni, A. Circumventing Scaling Relations in Oxygen Electrochemistry Using Metal-Organic Frameworks. The Journal of Physical Chemistry Letters 2020, 11, 10029–10036.DOI
  6. [7]Sliwa, M.; McGonegle, D.; Wehrenberg, C.; Bolme, C. A.; Heighway, P. G.; Higginbotham, A.; Lazicki, A.; Lee, H. J.; Nagler, B.; Park, H. S.; Rudd, R. E.; Suggit, M. J.; Swift, D.; Tavella, F.; Zepeda-Ruiz, L.; Remington, B. A.; Wark, J. S. Femtosecond X-Ray Diffraction Studies of the Reversal of the Microstructural Effects of Plastic Deformation during Shock Release of Tantalum. Phys. Rev. Lett. 2018, 120, 265502.DOI
  7. [8]Stöhr, M.; Medrano Sandonas, L.; Tkatchenko, A. Accurate Many-Body Repulsive Potentials for Density-Functional Tight Binding from Deep Tensor Neural Networks. The Journal of Physical Chemistry Letters 2020, 11, 6835–6843.
  8. [9]Dolgоnos, G.; Aradi, B.; Moreira, N. H.; Frauenheim, T. An Improved Self-Consistent-Charge Density-Functional Tight-Binding (SCC-DFTB) Set of Parameters for Simulation of Bulk and Molecular Systems Involving Titanium. J. Chem. Theory Comput. 2010, 6, 266.
  9. [10]Lourenço, M. P.; dos Santos, E. C.; Pettersson, L. G. M.; Duarte, H. A. Accurate SCC-DFTB Parametrization for Bulk Water. Journal of Chemical Theory and Computation 2020, 16, 1768–1778.DOI
  10. [11]Aradi, B.; Hourahine, B.; Frauenheim, T. DFTB+, a sparse matrix-based implementation of the DFTB method. J. Phys. Chem. A 2007, 111, 5678–5684.pubmed.ncbi.nlm.nih.gov/17567110
  11. [12]Kullgren, J.; Wolf, M. J.; Hermansson, K.; Köhler, C.; Aradi, B.; Fauentheim, T.; Broqvist, P. Self-Consistent-Charge Density-Functional Tight-Binding (SCC-DFTB) Parameters for Ceria in 0D to 3D. J. Phys. Chem. C 2017, 121, 4593–4607.DOI
  12. [13]Goldman, N.; Koziol, L.; Fried, L. E. Using force-matched potentials to improve the accuracy of density functional tight binding for reactive conditions. J. Chem. Theory Comput. 2015, 11, 4530–4535.
  13. [14]Goldman, N.; Srinivasan, S. G.; Hamel, S.; Fried, L. E.; Gaus, M.; Elstner, M. Determination of a density functional tight binding model with an extended basis set and three-body repulsion for carbon under extreme pressures and temperatures. J. Phys. Chem. C 2013, 117, 7885 – 7894.DOI
  14. [15]Söderlind, P.; Zhou, F.; Landa, A.; Klepeis, J. E. Phonon and magnetic structure in δ-plutonium from density functional theory. Scientific Reports 2015, 5, 15958.DOI
  15. [16]Cheng, B.; Engel, E. A.; Behler, J.; Dellago, C.; Ceriotti, M. Ab initio thermodynamics of liquid and solid water. Proc. Natl. Acad. Sci. U.S.A. 2019, 116, 1110–1115.
  16. [17]Lindsey, R. K.; Fried, L. E.; Goldman, N. ChIMES: A Force Matched Potential with Explicit Three-Body Interactions for Molten Carbon. J. Chem. Theory Comput. 2017, 13, 6222–6229.DOI
  17. [18]Lindsey, R. K.; Fried, L. E.; Goldman, N.; Bastea, S. Active learning for robust, high-complexity reactive atomistic simulations. J. Chem. Phys. 2020, 153, 134117.DOI
  18. [19]Koziol, L.; Fried, L. E.; Goldman, N. Using Force Matching To Determine Reactive Force Fields for Water under Extreme Thermodynamic Conditions. J. Chem. Theory Comput. 2017, 13, 135–146.DOI
  19. [20]Lindsey, R. K.; Fried, L. E.; Goldman, N. Application of the ChIMES Force Field to Nonreactive Molecular Systems: Water at Ambient Conditions. Journal of Chemical Theory and Computation 2019, 15, 436–447.DOI
  20. [21]Armstrong, M. R.; Lindsey, R. K.; Goldman, N.; Nielsen, M. H.; Stavrou, E.; Fried, L. E.; Zaug, J. M.; Bastea, S. Ultrafast shock synthesis of nanocarbon from a liquid precursor. Nature Communications 2020, 11, 353.DOI
  21. [22]Lindsey, R. K.; Goldman, N.; Fried, L. E.; Bastea, S. Many-body reactive force field development for carbon condensation in C/O systems under extreme conditions. J. Chem. Phys. 2020, 153, 054103.DOI
  22. [23]Pham, C. H.; Lindsey, R. K.; Fried, L. E.; Goldman, N. Calculation of the detonation state of HN₃ with quantum accuracy. J. Chem. Phys. 2021, 153, 224102.DOI
  23. [24]Tersoff, J. Empirical interatomic potential for carbon, with application to amorphous-carbon. Phys. Rev. Lett. 1988, 61, 2879.pubmed.ncbi.nlm.nih.gov/10039251
  24. [27]Kitabayashi, K.; Edalati, K.; Li, H.-W.; Akiba, E.; Horita, Z. Phase Transformations in MgH₂-TiH₂ Hydrogen Storage System by High-Pressure Torsion Process. Advanced Engineering Materials 2020, 22, 1900027.DOI
  25. [28]Shanavas, K. V.; Lindsay, L.; Parker, D. S. Electronic structure and electron-phonon coupling in TiH₂. Scientific Reports 2016, 6, 28102.DOI
  26. [29]Sheppard, D. A.; Paskevicius, M.; Humphries, T. D.; Felderhoff, M.; Capurso, G.; Bellosta von Colbe, J.; Dornheim, M.; Klassen, T.; Ward, P. A.; Teprovich, J. A.; Corgnale, C.; Zidan, R.; Grant, D. M.; Buckley, C. E. Metal hydrides for concentrating solar thermal power energy storage. Applied Physics A 2016, 122, 395.DOI
  27. [30]Kovalev, D. Y.; Ratnikov, V. K. P. V. I.; Ponomarev, V. I. Thermal Decomposition of TiH₂: A TRXRD Study. International Journal of Self Propagating High Temperature Synthesis 2010, 19, 253–257.DOI
  28. [31]Sandim, H. R. Z.; Morante, B. V.; Suzuki, P. A. Kinetics of Thermal Decomposition of Titanium Hydride Powder Using in situ High-temperature X-ray Diffraction (HTXRD). Materials Research 2005, 8, 293–297.DOI
  29. [32]Peng, Q.; Yang, B.; Liu, L.; Song, C.; Friedrich, B. Porous TiAl alloys fabricated by sintering of TiH2 and Al powder mixtures. Journal of Alloys and Compounds 2016, 656, 530–538.DOI
  30. [33]Yang, Y.; Zhao, F.; Huang, X.; Zhang, J.; Chen, X.; Yuan, Z.; Wang, Y.; Li, H.; Li, S. Reinforced combustion of the ZrH2-HMX-CMDB propellant: The critical role of hydrogen. Chemical Engineering Journal 2020, 402, 126275.DOI
  31. [34]Heo, T. W.; Colas, K. B.; Motta, A. T.; Chen, L.-Q. A phase-field model for hydride formation in polycrystalline metals: Application to delta-hydride in zirconium alloys. Acta Mater. 2019, 181, 262–277.DOI
  32. [35]Jing, P.; Liu, M.; Pu, Y.; Cui, Y.; Wang, Z.; Wang, J.; Q.Liu, Dependence of phase configurations, microstructures and magnetic properties of iron-nickel (Fe-Ni) alloy nanoribbons on deoxidization temperature in hydrogen. Sci. Rep. 2016, 6, 37701.DOI
  33. [36]Kresse, G.; Hafner, J. Ab initio molecular dynamics for liquid metals. Phys. Rev. B 1993, 47, 558–561.
  34. [37]Kresse, G.; Hafner, J. Ab initio molecular dynamics simulation of the liquid-metal-amorphous-semiconductor transition in germanium. Phys. Rev. B 1994, 49, 14251–14271.pubmed.ncbi.nlm.nih.gov/10010505
  35. [38]Kresse, G.; Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 1996, 54, 11169–11186.pubmed.ncbi.nlm.nih.gov/9984901
  36. [39]Blöchl, P. E. Projector augmented-wave method. Phys. Rev. B 1994, 50, 17953–17979.pubmed.ncbi.nlm.nih.gov/9976227
  37. [40]Kresse, G.; Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B 1999, 59, 1758–1775.DOI
  38. [41]Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 1996, 77, 3865–3868.pubmed.ncbi.nlm.nih.gov/10062328
  39. [42]Methfessel, M.; Paxton, A. T. High-precision sampling for Brillouin-zone integration in metals. Phys. Rev. B 1989, 40, 3616–3621.pubmed.ncbi.nlm.nih.gov/9992329
  40. [43]Mermin, N. D. Thermal properties of the Inhomogenous Electron Gas. Phys. Rev. 1965, 137, 1441–1443.DOI
  41. [44]Nosé, S. Molecular Physics 1984, 52, 255.
  42. [45]Hoover, W. G. Physical Review A 1985, 31, 1695.
  43. [46]Martyna, G. J.; Klein, M. L.; Tuckerman, M. Nosé-Hoover chains: The canonical ensemble via continuous dynamics. The Journal of Chemical Physics 1992, 97, 2635–2643.DOI
  44. [47]Elstner, M.; Porezag, D.; Jungnickel, G.; Elsner, J.; Haugk, M.; Frauenheim, T.; Suhai, S.; Seifert, G. Self-consistent-charge density-functional tight-binding method for simulations of complex materials properties. Phys. Rev. B 1998, 58, 7260–7268.DOI
  45. [48]Koskinen, P.; Mäkinen, V. Density-functional tight-binding for beginners. Comp. Mater. Sci. 2009, 47, 237–253.DOI
  46. [49]Gaus, M.; Cui, Q.; Elstner, M. DFTB3: Extension of the Self-Consistent-Charge Density-Functional Tight-Binding Method (SCC-DFTB). J. Chem. Theory Comput. 2011, 7, 931.pubmed.ncbi.nlm.nih.gov/23204947
  47. [50]Hourahine, B.; Aradi, B.; Blum, V.; Bonafè, F.; Buccheri, A.; Camacho, C.; Cevallos, C.; Deshaye, M. Y.; Dumitričă, T.; Dominguez, A.; Ehlert, S.; Elstner, M.; van der Heide, T.; Hermann, J.; Irle, S.; Kranz, J. J.; Köhler, C.; Kowalczyk, T.; Kubař, T.; Lee, I. S.; Lutske r, V.; Maurer, R. J.; Min, S. K.; Mitchell, I.; Negre, C.; Niehaus, T. A.; Niklasson, A. M. N.; Page, A. J.; Pecchia, A.; Penazzi, G.; Persson, M. P.; Řezáč, J.; Sánchez, C. G.; Sternberg, M.; Stöhr, M.; Stuckenberg, F.; Tkatchenko, A.; Yu, V. W.-z.; Frauenheim, T. DFTB+, a software package for efficient approximate density functional theory based atomistic simulations. The Journal of Chemical Physics 2020, 152, 124101.
  48. [51]Christensen, A. S.; Kubár, T.; Cui, Q.; Elstner, M. Semiempirical Quantum Mechanical Methods for Noncovalent Interactions for Chemical and Biochemical Applications. Chemical Reviews 2016, 116, 5301–5337.DOI
  49. [52]Porezag, D.; Frauenheim, T.; Köhler, T.; Seifert, G.; Kaschner, R. K. Construction of tight-binding-like potentials on the basis of density functional theory: Application to carbon. Phys. Rev. B 1995, 51, 12947.pubmed.ncbi.nlm.nih.gov/9978089
  50. [53]Gaus, M.; Goetz, A.; Elstner, M. Parametrization and Benchmark of DFTB3 for Organic Molecules. Journal of Chemical Theory and Computation 2012, 9, 338–354.DOI
  51. [54]Gaus, M.; Cui, Q.; Elstner, M. DFTB3: Extension of the Self-Consistent-Charge Density-Functional Tight-Binding Method (SCC-DFTB). J. Chem. Theory Comput. 2011, 7, 931–948.pubmed.ncbi.nlm.nih.gov/23204947
  52. [55]Vuong, V. Q.; Madridjejos, J. M. L.; Aradi, B.; Sumpter, B. G.; Metha, G. F.; Irle, S. Density-functional tight-binding for phosphine-stabilized nanoscale gold clusters. Chem. Sci. 2020, 11, 13113–13128.DOI
  53. [56]Goldman, N.; Aradi, B.; Lindsey, R. K.; Fried, L. E. Development of a Multicenter Density Functional Tight Binding Model for Plutonium Surface Hydriding. J. Chem. Theory Comput. 2018, 14, 2652–2660.DOI
  54. [57]Dantanarayana, V.; Nemetiaram, T.; Vong, D.; Anthony, J. E.; Troisi, A.; Cong, K. N.; Goldman, N.; Faller, R.; Moulé, A. J. Predictive Model of Charge Mobilities in Organic Semiconductor Small Molecules with Force-Matched Potentials. J. Chem. Theory Comput. 2020, 16, 3494–3503.DOI
  55. [58]Rappé, A. K.; Casewit, C. J.; Colwell, K. S.; III, W. A. G.; Skiff, W. M. UFF, a Full Periodic Table Force Field for Molecular Mechanics and Molecular Dynamics Simulations. J. Am. Chem. Soc. 1992, 114, 10024–10039.DOI
  56. [59]Grimme, S.; Antony, J.; Ehrlich, S.; Krieg, H. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. The Journal of Chemical Physics 2010, 132, 154104.DOI
  57. [60]Wang, Y.; Shepler, B. C.; Braams, B. J.; Bowman, J. M. Full-dimensional, ab initio potential energy and dipole moment surfaces for water. J. Chem. Phys. 2009, 131, 054511.DOI
  58. [61]Wang, Y.; Huang, X.; Shepler, B. C.; Braams, B. J.; Bowman, J. M. Flexible, ab initio potential, and dipole moment surfaces for water. I. Tests and applications for clusters up to the 22-mer. J. Chem. Phys. 2011, 134, 094509.DOI
  59. [62]Luschtinetz, R.; Frenzel, J.; Milek, T.; Seifert, G. Adsorption of phosphonic acid at the TiO₂ anatase (101) and rutile (110) surface. J. Phys. Chem. C 2009, 113, 5730–5740.DOI
  60. [63]Zheng, G.; Witek, H. A.; Bobadova-Parvanova, P.; Irle, S.; Musaëv, D. G.; Prabhakar, R.; Morokuma, K.; Lundberg, M.; Elstner, M.; Kohler, C.; Frauenheim, T. Parameter Calibration of Transition-Metal Elements for the Spin-Polarized Self-Consistent-Charge Density-Functional Tight-Binding (DFTB) Method: Sc, Ti, Fe, Co, and Ni. J. Chem. Theory Comput. 2007, 3, 1349–1367.DOI
  61. [64]Dolgonos, G.; Aradi, B.; Moreira, N. H.; Frauenheim, T. An Improved Self-Consistent-Charge Density-Functional Tight-Binding (SCC-DFTB) Set of Parameters for Simulation of Bulk and Molecular Systems Involving Titanium. J. Chem. Theory Comput. 2010, 6, 266–278.
  62. [65]Goyal, P.; Qian, H.-J.; Irle, S.; Lu, X.; Roston, D.; Mori, T.; Elstner, M.; Cui, Q. Molecular Simulation of Water and Hydration Effects in Different Environments: Challenges and Developments for DFTB Based Models. The Journal of Physical Chemistry B 2014, 118, 11007–11027.
  63. [67]Press, W.; Flannery, B. P.; Teukolsky, S. A.; Wetterling, W. T. Numerical Recipes; Cambridge University Press: Cambridge, 1989.
  64. [68]Efron, B.; Hastie, T.; Johnstone, I.; Tibshirani, R. Least angle regression. The Annals of statistics 2004, 32, 407–499.
  65. [69]Friedman, J.; Hastie, T.; Tibshirani, R. Regularization paths for generalized linear models via coordinate descent. Journal of statistical software 2010, 33, 1.pubmed.ncbi.nlm.nih.gov/20808728
  66. [70]Tibshirani, R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological) 1996, 58, 267–288.
  67. [71]Moser, D.; Bull, D. J.; Sato, T.; Noréus, D.; Kyoi, D.; Sakai, T.; Kitamura, N.; Yusa, H.; Taniguchi, T.; Kalisvaart, W. P.; Notten, P. Structure and stability of high pressure synthesized Mg-TM hydrides (TM = Ti, Zr, Hf, V, Nb and Ta) as possible new hydrogen rich hydrides for hydrogen storage. J. Mater. Chem. 2009, 19, 8150–8161.DOI
  68. [72]Bodrog, Z.; Aradi, B. Possible improvements to the self-consistent-charges density-functional tight-binding method within the second order. Phys. Status Solidi B 2012, 259, 259–269.DOI
  69. [73]Goldman, N. Multi-center semi-empirical quantum models for carbon under extreme thermodynamic conditions. Chem. Phys. Lett. 2015, 622, 128–136.
  70. [74]Podeszwa, R.; Jankiewicz, W.; Krzúś, M.; Witek, H. A. Correcting long-range electrostatics in DFTB. J. Chem. Phys. 2019, 150, 234110.DOI
  71. [75]Wyckoff, R. W. G. Crystal Structures, 1st ed.; John Wiley, 1963.
  72. [76]Yoshimoto, Y. Extended multicanonical method combined with thermodynamically optimized potential: Application to the liquid-crystal transition of silicon. The Journal of Chemical Physics 2006, 125, 184103.pubmed.ncbi.nlm.nih.gov/17115734

Paper details

vlnc:W_4mdn4h6w
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents