Real-Time Time-Dependent Electronic Structure Theory

I have a new review on (and titled) real-time time-dependent electronic structure theory out now, which has been one of my active research areas over the past five years or so. Like other electronic structure methods, real-time time-dependent (RT-TD) methods seeks to understand molecules through a quantum treatment of the motions of their electrons. What sets RT-TD theories and methods apart is that they explore how electrons evolve in time, especially as they respond to things like lasers, magnetic fields, X-rays, and so on. Real time methods look at molecular responses to these things explicitly in time, and gives an intuitive and dynamic view of how molecules behave under all sorts of experimental conditions. From this, we can predict and explain how certain molecules make better candidates for solar energy conversion, molecular electronics, or nanosensors, etc.

The truth is that there is a wealth of information that can be obtained by real-time time-dependent methods. Because of this, RT-TD methods are often criticized as inefficient and overkill for many predictive applications. In some cases this is true, for example when computing absorption spectra of small organic molecules (linear response methods are usually a better choice). However, to dismiss all RT-TD methods is a mistake, as the methods comprise a general technique for studying extremely complex phenomena. RT-TD methods allow for complete control over the number and strength of interacting external perturbations (multiple intense lasers, for example) in the study of molecular responses. In the review, we address many of these unique applications, ranging from non-equilibrium solvent dynamics to dynamic hyperpolarizability to the emerging real-time quantum electrodynamics (QED), where even the photon field is quantized.

The article was extremely fun to write, and I hope you find something useful and interesting in it. RT-TD comprises a broad field with extremely talented scientists working on its development. You can find the article here.

The Magnus Expansion

PDF version

The goal of all real-time electronic dynamics methods is to solve the time-dependent Schrödinger equation (TDSE)

where \(H(t)\) is the time-dependent Hamiltonian and \(\psi(t)\) is the time-dependent wave function. The goal of the Magnus expansion is to find a general solution for the time-dependent wave function in the case where \(H\) is time-dependent, and, more crucially, when \(H\) does not commute with itself at different times, e.g. when \(\left[H(t_1),H(t_2)\right] \neq 0\). In the following we will follow closely the notation of Blanes, et al..

First, for simplicity we redefine \(\tilde{H}(t) \equiv \frac{-i}{\hbar}H(t)\) and introduce a scalar \(\lambda = 1\) as a bookkeeping device, so that

At the heart of the Magnus expansion is the idea of solving the TDSE by using the quantum propagator \(U(t,t_0)\) that connects wave functions at different times, e.g. \(\psi(t) = U(t,t_0)\psi(t_0)\) Furthermore, the Magnus expansion assumes that \(U(t,t_0)\) can be represented as an exponential, \(U(t,t_0) = \text{exp} \left( \Omega(t,t_0) \right)\) This yields the modified TDSE

Now, for scalar \(H\) and \(U\), the above has a simple solution, namely

However, if \(H\) and \(U\) are matrices this is not necessarily true. In other words, for a given matrix \(A\) the following expression does not necessarily hold:

because the matrix \(A\) and its derivatives do not necessarily commute. Instead, Magnus proved that in general \(\Omega (t, t_0)\) satisfies

where \(B_k\) are the Bernoulli numbers. This equation may be solved by integration, and iterative substitution of \(\Omega(t,t_0)\). While it may appear that we are worse off than when we started, collecting like powers of \(\lambda\) (and setting \(\lambda = 1\)) allows us to obtain a power-series expansion for \(\Omega(t,t_0)\),

This is the Magnus expansion, and here we have given up to the third-order terms. We have also made the notational simplification that \(\tilde{H}_k = \tilde{H}(t_k)\). This is the basis for nearly all numerical methods to integrate the many-body TDSE in molecular physics. Each subsequent order in the Magnus expansion is a correction that accounts for the proper time-ordering of the Hamiltonian.

The Magnus expansion immediately suggests a route to many numerical integrators. The simplest would be to approximate the first term by

leading to a forward-Euler-like time integrator of

which we can re-write as

where subscript \(k\) gives the node of the time-step stencil. This gives a first-order method with error \(\mathcal{O}(\Delta t)\). A more accurate second-order method can be constructed by approximating the first term in the Magnus expansion by the midpoint rule, leading to an \(\mathcal{O}({\Delta t}^2)\) time integrator

Modifying the stencil to eliminate the need to evaluate the Hamiltonian at fractional time steps (e.g. change time step to \(2 \Delta t\)) leads to the modified midpoint unitary transformation (MMUT) method

which is a leapfrog-type unitary integrator. Note that the midpoint method assumes \(\tilde{H}\) is linear over its time interval, and the higher order terms (containing the commutators) in this approximation go to zero. There are many other types of integrators based off the Magnus expansion that can be found in the literature. The key point for all of these integrators is that they are symplectic, meaning they preserve phase-space relationships. This has the practical effect of conserving energy (within some error bound) in long-time dynamics, whereas non-symplectic methods such as Runge-Kutta will experience energetic “drift” over long times. A final note: in each of these schemes it is necessary to evaluate the exponential of the Hamiltonian. In real-time methods, this requires computing a matrix exponential. This is not a trivial task, and, aside from the construction of the Hamiltonian itself, is often the most expensive step in the numerical solution of the TDSE. However, many elegant solutions to the construction of the matrix exponential can be found in the literature.

References

Blanes, S., Casas, F., Oteo, J.A. and Ros, J., 2010. A pedagogical approach to the Magnus expansion. European Journal of Physics, 31(4), p.907.

Magnus, W., 1954. On the exponential solution of differential equations for a linear operator. Communications on pure and applied mathematics, 7(4), pp.649-673.

A (hopefully) gentle guide to the computer implementation of molecular integrals over Gaussian basis functions.

Note: You can find the following integral routines implemented in a working Hartree-Fock code here.

Special thank you to Joy Alamgir for pointing out several corrections to this post (and PDF)!

PDF version

In quantum chemistry, we are often interested in evaluating integrals over Gaussian basis functions. Here I am going to take much of the theory for granted and talk about how one might actually implement the integrals for quantum chemistry by using recursion relationships with Hermite Gaussians. Our goal is to have clean, readable code. I’m writing this for the people are comfortable with the mathematics behind the Gaussian integrals, but want to see a readable computer implementation. I’ll talk a bit about some computational considerations at the end, but my goal is to convert equations to code. In fact, I’ve tried to structure the equations and the code in such a way that the two look very similar.

For a very good overview of integral evaluation, please see:

Helgaker, Trygve, and Peter R. Taylor. “Gaussian basis sets and molecular integrals.” Modern Electronic Structure (1995).

I will try and follow the notation used in the above reference.

Mathematical preliminaries

Let’s start with some of the basics. First, we have our 3D Gaussian functions

with orbital exponent \(\alpha\), electronic coordinates \(\mathbf{r}\), origin \(\mathbf{A}\), and

also, \(i,j,k\) are the angular quantum numbers (e.g. \(i=0\) is \(s\)-type, \(i=1\) is \(p\) type, etc.) Cartesian Gaussians are separable in 3D along \(x,y,z\) so that

with the 1D Gaussian

So far, so good. Let’s consider the overlap integral of two 1D Gaussians, \(a\) and \(b\)

where we used the Gaussian product theorem so that

When using Hermite Gaussians, we can express \(S_{ab}\) as

where \(E_{t}^{ij}\) are expansion coefficients (to be determined recursively) and \(\Lambda_t\) is the Hermite Gaussian overlap of two Gaussians \(a\) and \(b\). It has a simple expression that kills the sum via the Kronecker delta \(\delta_{t0}\). It can be shown that the expansion coefficients can be defined using the following recursive definitions

The first equation gives us a way to reduce the index \(j\) and the second gives us a way to reduce index \(i\) so that we can get to the third equation, which is our base case. The last equation tells us what to do if we go out of index bounds.

Overlap integrals

The first thing we need to do is implement a function E which returns our expansion coefficients \(E_t^{ij}\). Aside from angular momentum \(i\) and \(j\) from the Gaussian functions, we also need the distance between Gaussians \(Q_x\) and the orbital exponent coefficients \(\alpha\) and \(\beta\) as inputs.

def E(i,j,t,Qx,a,b):
    ''' Recursive definition of Hermite Gaussian coefficients.
        Returns a float.
        a: orbital exponent on Gaussian 'a' (e.g. alpha in the text)
        b: orbital exponent on Gaussian 'b' (e.g. beta in the text)
        i,j: orbital angular momentum number on Gaussian 'a' and 'b'
        t: number nodes in Hermite (depends on type of integral, 
           e.g. always zero for overlap integrals)
        Qx: distance between origins of Gaussian 'a' and 'b'
    '''
    p = a + b
    q = a*b/p
    if (t < 0) or (t > (i + j)):
        # out of bounds for t  
        return 0.0
    elif i == j == t == 0:
        # base case
        return np.exp(-q*Qx*Qx) # K_AB
    elif j == 0:
        # decrement index i
        return (1/(2*p))*E(i-1,j,t-1,Qx,a,b) - \
               (q*Qx/a)*E(i-1,j,t,Qx,a,b)    + \
               (t+1)*E(i-1,j,t+1,Qx,a,b)
    else:
        # decrement index j
        return (1/(2*p))*E(i,j-1,t-1,Qx,a,b) + \
               (q*Qx/b)*E(i,j-1,t,Qx,a,b)    + \
               (t+1)*E(i,j-1,t+1,Qx,a,b)

This is simple enough! So for a 1D overlap between two Gaussians we would just need to evaluate \(E_0^{ij}\) and multiply it by \(\sqrt{\frac{\pi}{p}}\). Overlap integrals in 3D are just a product of the \(x,y,z\) 1D overlaps. We could imagine a 3D overlap function like so

import numpy as np

def overlap(a,lmn1,A,b,lmn2,B):
    ''' Evaluates overlap integral between two Gaussians
        Returns a float.
        a:    orbital exponent on Gaussian 'a' (e.g. alpha in the text)
        b:    orbital exponent on Gaussian 'b' (e.g. beta in the text)
        lmn1: int tuple containing orbital angular momentum (e.g. (1,0,0))
              for Gaussian 'a'
        lmn2: int tuple containing orbital angular momentum for Gaussian 'b'
        A:    list containing origin of Gaussian 'a', e.g. [1.0, 2.0, 0.0]
        B:    list containing origin of Gaussian 'b'
    '''
    l1,m1,n1 = lmn1 # shell angular momentum on Gaussian 'a'
    l2,m2,n2 = lmn2 # shell angular momentum on Gaussian 'b'
    S1 = E(l1,l2,0,A[0]-B[0],a,b) # X
    S2 = E(m1,m2,0,A[1]-B[1],a,b) # Y
    S3 = E(n1,n2,0,A[2]-B[2],a,b) # Z
    return S1*S2*S3*np.power(np.pi/(a+b),1.5) 

Note that we are using the NumPy package in order to take advantage of the definitions of \(\pi\) and the fractional power to the \(3/2\). The above two functions overlap and E are enough to get us the overlap between two Gaussian functions (primitives), but most basis functions are contracted, meaning they are the sum of multiple Gaussian primitives. It is not too difficult to account for this, and we can finally wrap up our evaluation of overlap integrals with a function S(a,b) which returns the overlap integral between two contracted Gaussian functions.

def S(a,b):
    '''Evaluates overlap between two contracted Gaussians
       Returns float.
       Arguments:
       a: contracted Gaussian 'a', BasisFunction object
       b: contracted Gaussian 'b', BasisFunction object
    '''
    s = 0.0
    for ia, ca in enumerate(a.coefs):
        for ib, cb in enumerate(b.coefs):
            s += a.norm[ia]*b.norm[ib]*ca*cb*\
                     overlap(a.exps[ia],a.shell,a.origin,
                     b.exps[ib],b.shell,b.origin)
    return s

Basically, this is just a sum over primitive overlaps, weighted by normalization and coefficient. A word is in order for the arguments, however. In order to keep the number of arguments we have to pass into our functions, we have created BasisFunction objects that contain all the relevant data for the basis function, including exponents, normalization, etc. A BasisFunction class looks like

from scipy.special import factorial2 as fact2

class BasisFunction(object):
    ''' A class that contains all our basis function data
        Attributes:
        origin: array/list containing the coordinates of the Gaussian origin
        shell:  tuple of angular momentum
        exps:   list of primitive Gaussian exponents
        coefs:  list of primitive Gaussian coefficients
        norm:   list of normalization factors for Gaussian primitives
    '''
    def __init__(self,origin=[0.0,0.0,0.0],shell=(0,0,0),exps=[],coefs=[]):
        self.origin = np.asarray(origin)
        self.shell = shell
        self.exps  = exps
        self.coefs = coefs
        self.norm = None
        self.normalize()

    def normalize(self):
        ''' Routine to normalize the basis functions, in case they
            do not integrate to unity.
        '''
        l,m,n = self.shell
        L = l+m+n
        # self.norm is a list of length equal to number primitives
        # normalize primitives first (PGBFs)
        self.norm = np.sqrt(np.power(2,2*(l+m+n)+1.5)*
                        np.power(self.exps,l+m+n+1.5)/
                        fact2(2*l-1)/fact2(2*m-1)/
                        fact2(2*n-1)/np.power(np.pi,1.5))

        # now normalize the contracted basis functions (CGBFs)
        # Eq. 1.44 of Valeev integral whitepaper
        prefactor = np.power(np.pi,1.5)*\
            fact2(2*l - 1)*fact2(2*m - 1)*fact2(2*n - 1)/np.power(2.0,L)

        N = 0.0
        num_exps = len(self.exps)
        for ia in range(num_exps):
            for ib in range(num_exps):
                N += self.norm[ia]*self.norm[ib]*self.coefs[ia]*self.coefs[ib]/\
                         np.power(self.exps[ia] + self.exps[ib],L+1.5)

        N *= prefactor
        N = np.power(N,-0.5)
        for ia in range(num_exps):
            self.coefs[ia] *= N

So, for example if we had a STO-3G Hydrogen 1s at origin (1.0, 2.0, 3.0), we could create a basis function object for it like so

myOrigin = [1.0, 2.0, 3.0]
myShell  = (0,0,0) # p-orbitals would be (1,0,0) or (0,1,0) or (0,0,1), etc.
myExps   = [3.42525091, 0.62391373, 0.16885540] 
myCoefs  = [0.15432897, 0.53532814, 0.44463454]
a = BasisFunction(origin=myOrigin,shell=myShell,exps=myExps,coefs=myCoefs)

Where we used the EMSL STO-3G definition

!  STO-3G  EMSL  Basis Set Exchange Library 
! Elements                             References
! --------                             ----------
!  H - Ne: W.J. Hehre, R.F. Stewart and J.A. Pople, J. Chem. Phys. 2657 (1969).

****
H     0 
S   3   1.00
      3.42525091             0.15432897       
      0.62391373             0.53532814       
      0.16885540             0.44463454       
****

So doing S(a,a) = 1.0, since the overlap of a basis function with itself (appropriately normalized) is one.

Kinetic energy integrals

Having finished the overlap integrals, we move on to the kinetic integrals. The kinetic energy integrals can be written in terms of overlap integrals

where

For a 3D primitive, we can form a kinetic function analogous to overlap,

def kinetic(a,lmn1,A,b,lmn2,B):
    ''' Evaluates kinetic energy integral between two Gaussians
        Returns a float.
        a:    orbital exponent on Gaussian 'a' (e.g. alpha in the text)
        b:    orbital exponent on Gaussian 'b' (e.g. beta in the text)
        lmn1: int tuple containing orbital angular momentum (e.g. (1,0,0))
              for Gaussian 'a'
        lmn2: int tuple containing orbital angular momentum for Gaussian 'b'
        A:    list containing origin of Gaussian 'a', e.g. [1.0, 2.0, 0.0]
        B:    list containing origin of Gaussian 'b'
    '''
    l1,m1,n1 = lmn1
    l2,m2,n2 = lmn2
    term0 = b*(2*(l2+m2+n2)+3)*\
                            overlap(a,(l1,m1,n1),A,b,(l2,m2,n2),B)
    term1 = -2*np.power(b,2)*\
                           (overlap(a,(l1,m1,n1),A,b,(l2+2,m2,n2),B) +
                            overlap(a,(l1,m1,n1),A,b,(l2,m2+2,n2),B) +
                            overlap(a,(l1,m1,n1),A,b,(l2,m2,n2+2),B))
    term2 = -0.5*(l2*(l2-1)*overlap(a,(l1,m1,n1),A,b,(l2-2,m2,n2),B) +
                  m2*(m2-1)*overlap(a,(l1,m1,n1),A,b,(l2,m2-2,n2),B) +
                  n2*(n2-1)*overlap(a,(l1,m1,n1),A,b,(l2,m2,n2-2),B))
    return term0+term1+term2

and for contracted Gaussians we make our final function T(a,b)

def T(a,b):
    '''Evaluates kinetic energy between two contracted Gaussians
       Returns float.
       Arguments:
       a: contracted Gaussian 'a', BasisFunction object
       b: contracted Gaussian 'b', BasisFunction object
    '''
    t = 0.0
    for ia, ca in enumerate(a.coefs):
        for ib, cb in enumerate(b.coefs):
            t += a.norm[ia]*b.norm[ib]*ca*cb*\
                     kinetic(a.exps[ia],a.shell,a.origin,\
                     b.exps[ib],b.shell,b.origin)
    return t

Nuclear attraction integrals

The last one-body integral I want to consider here is the nuclear attraction integrals. These differ from the overlap and kinetic energy integrals in that the nuclear attraction operator \(1/r_C\) is Coulombic, meaning we cannot easily factor the integral into Cartesian components \(x,y,z\).

To evaluate these integrals, we need to set up an auxiliary Hermite Coulomb integral \(R^n_{tuv}(p,\mathbf{P},\mathbf{C})\) that handles the Coulomb interaction between a Gaussian charge distribution centered at \(\mathbf{P}\) and a nuclei centered at \(\mathbf{C}\). The Hermite Coulomb integral, like its counterpart \(E_t^{ij}\), is defined recursively:

where \(F_n(T)\) is the Boys function

which is a special case of the Kummer confluent hypergeometric function, \(_1F_1(a,b,x)\)

which is convenient for us, since SciPy has an implementation of \(_1F_1\) as a part of scipy.special. So for R we can code up the recursion like so

def R(t,u,v,n,p,PCx,PCy,PCz,RPC):
    ''' Returns the Coulomb auxiliary Hermite integrals 
        Returns a float.
        Arguments:
        t,u,v:   order of Coulomb Hermite derivative in x,y,z
                 (see defs in Helgaker and Taylor)
        n:       order of Boys function 
        PCx,y,z: Cartesian vector distance between Gaussian 
                 composite center P and nuclear center C
        RPC:     Distance between P and C
    '''
    T = p*RPC*RPC
    val = 0.0
    if t == u == v == 0:
        val += np.power(-2*p,n)*boys(n,T)
    elif t == u == 0:
        if v > 1:
            val += (v-1)*R(t,u,v-2,n+1,p,PCx,PCy,PCz,RPC)
        val += PCz*R(t,u,v-1,n+1,p,PCx,PCy,PCz,RPC)
    elif t == 0:
        if u > 1:
            val += (u-1)*R(t,u-2,v,n+1,p,PCx,PCy,PCz,RPC)
        val += PCy*R(t,u-1,v,n+1,p,PCx,PCy,PCz,RPC)
    else:
        if t > 1:
            val += (t-1)*R(t-2,u,v,n+1,p,PCx,PCy,PCz,RPC)
        val += PCx*R(t-1,u,v,n+1,p,PCx,PCy,PCz,RPC)
    return val

and we can define our boys(n,T) function as

from scipy.special import hyp1f1

def boys(n,T):
    return hyp1f1(n+0.5,n+1.5,-T)/(2.0*n+1.0) 

There are other definitions of the Boys function of course, in case you do not want to use the SciPy built-in. Note that R requires knowledge of the composite center \(\mathbf{P}\) from two Gaussians centered at \(\mathbf{A}\) and \(\mathbf{B}\). We can determine \(\mathbf{P}\) using the Gaussian product center rule

which is very simply coded up as

def gaussian_product_center(a,A,b,B):
    return (a*A+b*B)/(a+b)

Now that we have a the Coulomb auxiliary Hermite integrals \(R^{n}_{tuv}\), we can form the nuclear attraction integrals with respect to a given nucleus centered at \(\mathbf{C}\), \(V_{ab}(C)\), via the expression

def nuclear_attraction(a,lmn1,A,b,lmn2,B,C):
    ''' Evaluates kinetic energy integral between two Gaussians
         Returns a float.
         a:    orbital exponent on Gaussian 'a' (e.g. alpha in the text)
         b:    orbital exponent on Gaussian 'b' (e.g. beta in the text)
         lmn1: int tuple containing orbital angular momentum (e.g. (1,0,0))
               for Gaussian 'a'
         lmn2: int tuple containing orbital angular momentum for Gaussian 'b'
         A:    list containing origin of Gaussian 'a', e.g. [1.0, 2.0, 0.0]
         B:    list containing origin of Gaussian 'b'
         C:    list containing origin of nuclear center 'C'
    '''
    l1,m1,n1 = lmn1 
    l2,m2,n2 = lmn2
    p = a + b
    P = gaussian_product_center(a,A,b,B) # Gaussian composite center
    RPC = np.linalg.norm(P-C)

    val = 0.0
    for t in range(l1+l2+1):
        for u in range(m1+m2+1):
            for v in range(n1+n2+1):
                val += E(l1,l2,t,A[0]-B[0],a,b) * \
                       E(m1,m2,u,A[1]-B[1],a,b) * \
                       E(n1,n2,v,A[2]-B[2],a,b) * \
                       R(t,u,v,0,p,P[0]-C[0],P[1]-C[1],P[2]-C[2],RPC)
    val *= 2*np.pi/p 
    return val

And, just like all the other routines, we can wrap it up to treat contracted Gaussians like so:

def V(a,b,C):
    '''Evaluates overlap between two contracted Gaussians
       Returns float.
       Arguments:
       a: contracted Gaussian 'a', BasisFunction object
       b: contracted Gaussian 'b', BasisFunction object
       C: center of nucleus
    '''
    v = 0.0
    for ia, ca in enumerate(a.coefs):
        for ib, cb in enumerate(b.coefs):
            v += a.norm[ia]*b.norm[ib]*ca*cb*\
                     nuclear_attraction(a.exps[ia],a.shell,a.origin,
                     b.exps[ib],b.shell,b.origin,C)
    return v

Important: Note that this is the nuclear repulsion integral contribution from an atom centered at \(\mathbf{C}\). To get the full nuclear attraction contribution, you must sum over all the nuclei, as well as scale each term by the appropriate nuclear charge!

Two electron repulsion integrals

We are done with the necessary one-body integrals (for a basic Hartree-Fock energy code, at least) and are ready to move on to the two-body terms: the electron-electron repulsion integrals. Thankfully, much of the work has been done for us on account of the nuclear attraction one-body integrals.

In terms of Hermite integrals, to evaluate the two electron repulsion terms, we must evaluate the summation

which looks terrible and it is. However, recalling that \(p = \alpha + \beta\) letting \(q = \gamma + \delta\) (that is, the Gaussian exponents on \(a\) and \(b\), and \(c\) and \(d\)), we can write the equation in a similar form to the nuclear attraction integrals

def electron_repulsion(a,lmn1,A,b,lmn2,B,c,lmn3,C,d,lmn4,D):
    ''' Evaluates kinetic energy integral between two Gaussians
        Returns a float.
        a,b,c,d:   orbital exponent on Gaussian 'a','b','c','d'
        lmn1,lmn2
        lmn3,lmn4: int tuple containing orbital angular momentum
                   for Gaussian 'a','b','c','d', respectively
        A,B,C,D:   list containing origin of Gaussian 'a','b','c','d'
    '''
    l1,m1,n1 = lmn1
    l2,m2,n2 = lmn2
    l3,m3,n3 = lmn3
    l4,m4,n4 = lmn4
    p = a+b # composite exponent for P (from Gaussians 'a' and 'b')
    q = c+d # composite exponent for Q (from Gaussians 'c' and 'd')
    alpha = p*q/(p+q)
    P = gaussian_product_center(a,A,b,B) # A and B composite center
    Q = gaussian_product_center(c,C,d,D) # C and D composite center
    RPQ = np.linalg.norm(P-Q)

    val = 0.0
    for t in range(l1+l2+1):
        for u in range(m1+m2+1):
            for v in range(n1+n2+1):
                for tau in range(l3+l4+1):
                    for nu in range(m3+m4+1):
                        for phi in range(n3+n4+1):
                            val += E(l1,l2,t,A[0]-B[0],a,b) * \
                                   E(m1,m2,u,A[1]-B[1],a,b) * \
                                   E(n1,n2,v,A[2]-B[2],a,b) * \
                                   E(l3,l4,tau,C[0]-D[0],c,d) * \
                                   E(m3,m4,nu ,C[1]-D[1],c,d) * \
                                   E(n3,n4,phi,C[2]-D[2],c,d) * \
                                   np.power(-1,tau+nu+phi) * \
                                   R(t+tau,u+nu,v+phi,0,\
                                       alpha,P[0]-Q[0],P[1]-Q[1],P[2]-Q[2],RPQ)

    val *= 2*np.power(np.pi,2.5)/(p*q*np.sqrt(p+q))
    return val

And, for completeness’ sake, we wrap the above to handle contracted Gaussians

def ERI(a,b,c,d):
    '''Evaluates overlap between two contracted Gaussians
        Returns float.
        Arguments:
        a: contracted Gaussian 'a', BasisFunction object
        b: contracted Gaussian 'b', BasisFunction object
        c: contracted Gaussian 'b', BasisFunction object
        d: contracted Gaussian 'b', BasisFunction object
    '''
    eri = 0.0
    for ja, ca in enumerate(a.coefs):
        for jb, cb in enumerate(b.coefs):
            for jc, cc in enumerate(c.coefs):
                for jd, cd in enumerate(d.coefs):
                    eri += a.norm[ja]*b.norm[jb]*c.norm[jc]*d.norm[jd]*\
                             ca*cb*cc*cd*\
                             electron_repulsion(a.exps[ja],a.shell,a.origin,\
                                                b.exps[jb],b.shell,b.origin,\
                                                c.exps[jc],c.shell,c.origin,\
                                                d.exps[jd],d.shell,d.origin)
    return eri

And there you have it! All the integrals necessary for a Hartree-Fock SCF code.

Computational efficiency considerations

Our goal here has been to eliminate some of the confusion when it comes to connecting mathematics to actual computer code. So the code that we have shown is hopefully clear and looks nearly identical to the mathematical equations they are supposed to represent. This is one reason we chose Python as the code of choice to implement the integrals. It emphasizes readability.

If you use the code as is, you’ll find that you can only really handle small systems. To that end, I’ll give a few ideas on how to improve the integral code to actually be usable.

First, I would recommend not using pure Python. The problem is that we have some pretty deep loops in the code, and nested for loops will kill your speed if you insist on sticking with Python. Now, you can code in another language, but I would suggest rewriting some of the lower level routines with Cython (http://www.cython.org). Cython statically compiles your code to eliminate many of the Python calls that slow your loops down. In my experience, you will get several orders of magnitude speed up. This brings me to my second point. One of the problems with Python is that you have the global interpreter lock (GIL) which basically means you cannot run things in parallel. Many of the integral evaluation routines would do well if you could split the work up over multiple CPUs. If you rewrite some of the low level routines such that they do not make Python calls anymore, you can turn off the GIL and wrap the code in OpenMP directives, or even use Cython’s cython.parallel module. This will take some thought, but can definitely be done. Furthermore, removing the explicit recursion in the E and R functions and making them iterative would go a long way to speed up the code.

A couple other thoughts: be sure to exploit the permutational symmetry of the integrals. The two electron repulsion integrals, for example, can be done in \(1/8\) of the time just by exploiting these symmetries, which are unrelated to point group. Also, you can exploit many of the integral screening routines, since many of the two electron integrals are effectively zero. There are a lot of tricks out there in the literature, go check them out!

Quantum Dynamics Movies

One of the cool things about working with real-time electronic dynamics is that you can make awesome videos.

I recently made a video of a resonant electronic excitation of hydrogen peroxide.

You can check out the video by clicking the picture below.

The video was made using the real-time TDHF module in Chronus Quantum. We visualize how the electronic density changes over time.

For this simple STO-3G model, hydrogen peroxide strongly absorbs at 18.0 eV. So we apply an oscillating electric field at the same frequency and the electronic density goes nuts. You can do this for any frequency of light, however.

In the future, we can look at more realistic systems with real time TDDFT.

It’s important because this type of phenomena underlies our ability to make better solar cells and faster computer processing devices.

Easy transparent SSH tunnels

Most of my development work is done on remote computing clusters, which are usually firewalled.

This means I can’t log in directly when I work from home. I have to log into a network machine, and then log in to my development machine from there.

It gets kind of tedious to keep doing something along the lines of

home:>$ ssh network-host
network-host:>$ ssh development-machine
development-machine:>$

It also makes it a pain to copy files from home to the remote machine (and vice versa!).

However, you can easily set up a transparent tunnel through your network-host that lets you (effectively) log on to your development machine as if it weren’t even firewalled.

Just open up (or create) ~/.ssh/config:

Add the following lines (replacing network-host and development-machine with the names of your computers). You can name them whatever you want. Here I named each machine network and devel.

# Non-firewalled computer
Host network 
HostName network-machine

# Firewalled computer
Host devel
HostName development-machine 
ProxyCommand ssh -o 'ForwardAgent yes' network 'ssh-add && nc %h %p'

What this does is log you into the non-firewalled network-host, then establish a connection to the firewalled computer through netcat (nc in this case). You never actually see the network host, it acts on your behalf, invisibly in the background.

So, from your home computer, you can login directly to the development machine:

home:>$ ssh devel 
development-machine:>$

You can also scp directly from your development machine now, too.

home:>$ scp devel:~/path/to/files .
files                           100% 4084     4.0KB/s   00:00 
home:>$