“ENERGY MINIMIZATION TECHNIQUES”

BY

DR . SANTOSH V. GANDHI

AISSMS COLLEGE OF PHARMACY

For all except the very simplest systems the potential energy is
a complicated, multidimentional function of the coordinates.
For e.g – Energy of a conformation of ethane is a function of the
18 internal co-ordinates or 24 Cartesian co-ordinates that are
required to completely specify the structure.
As we have seen in Molecular Mechanics, the energy varies
with the co-ordinates is usually referred as potential energy
surface (Sometimes called as hyperspace)
For a system with N atoms the energy thus is a function of 3N-6
internal co-ordinates or 3N Cartesian co-ordinates. It is
therefore impossible to visualize the entire energy surface
except for some simple cases where the energy is a function of
one or two co-ordinates.
Sometimes we may visualize just a part of the energy
surface.
For e.g. Suppose we take an extended conformation of
pentane and rotate the two central carbon-carbon bonds so
that tortional angles vary from 0 to 360 degree, calculating
the energy of each structure generated.
The energy in this case is a function of just two variables and
can be plotted as contour diagram or isometric plot as shown
in figure
In Molecular Modeling, we are interested in minimum points on the energy surface.

Minimum energy arrangements of the atoms corresponds to stable states of the

system: any movement away from a minimum gives a configuration with a higher
energy.

There may be a very large number of minima on the energy surface. The minimum
with the very lowest energy is known as the Global Energy Minimum
 Saddle Point

Local Minimum
Global Minimum

Both Minima and Saddle points are stationary points on the energy surface, where
the first derivative of the energy function is zero with respect to all co-ordinates
Energy minimization

Energy minimization methods can precisely locate minimum

energy conformations by mathematically "homing in" on the
energy function minima (one at a time).

 The goal of energy minimization is to find a route

(consisting of variation of the intramolecular degrees of
freedom) from an initial conformation to the nearest minimum
energy conformation using the smallest number of calculations
possible.

All degrees of freedom in a molecule (3*number of atoms-5)

are usually included in the calculations, although most
programs allow you to geometrically tether (restrain) or fix
(constrain) specific atoms.
The relationship between an initial and minimum energy conformation is
shown in the following hypothetical energy surface:
The most direct route is a straight line in conformational hyperspace, but
minimization is rarely that efficient
Minimization algorithms are designed to head down-hill toward the nearest
minimum.

Remote minima (separated from the initial conformation by an energy barrier) are
not detected by energy minimization because this would require some period of up-
hill movement.

Energy minimization algorithms measure the energy surface along a series of

incremental steps to determine a down-hill direction:
 The local shape of the energy surface around a given conformation en route to a minimum
is often assumed to be quadratic so as to simplify the mathematics. This is illustrated for a
one-dimensional slice through an energy surface:

The quadratic energy function approximation is given by a Taylor series

where "P" is the current point and "x" an arbitrary point on the energy surface, "b" is the gradient
at "P", and "A" is the "Hessian Matrix" (the second partial derivatives) at "P." "A" and "b" can
be viewed as parameters that fit the idealized quadratic form to the actual energy surface. The
"A" term can improve the fit, allowing greater look-ahead (i.e. larger step sizes) toward a
minimum, but at the cost of greatly increased computational requirements.
An energy minimum can be characterized by a small change in energy between steps
and/or by a zero gradient of the energy function.

 The gradient usually provides the best indication for completion of minimization
(convergence) because very small changes in the gradient are much more significant than
small changes in the energy function itself.

 Precise convergence can, thus, be reached with a non-zero gradient. Under

some pathological situations, the gradient test can fail, and the energy test is then
used.

Energy minimization programs often limit the calculations to a set computer time
and/or a set number of steps/iterations by default. Ordinarily, you should
substantially increase any such limits.
The major difference between energy minimization methods lies in their approach to
calculating and applying the "b" term, whether the "A" term is included, and if so,
how "A" is calculated.

The major minimization techniques can be summarized as follows:

1. Calculate "b" at each step, but don't keep track of "b's" from previous steps.
2. Calculate "b" at each step, while keeping track of "b's" from previous steps.
3. Calculate "b" and "A" (or a first-derivative approximation of "A") at each step,
while keeping track of this information from previous steps.

• Each of these approaches has tradeoffs between computational

requirements/efficiency and avoidance of mathematical problems associated with
an actual energy surface compared to the quadratic model.

• The primary issues are:

1. Tolerance to poorly behaved regions of an energy function (e.g. bad interatomic
contacts, poorly defined minima).
2. Stability/reliability/precision.
Problematic kinds of energy surfaces are associated with small changes in energy
and gradient en route to a minimum, or even around a minimum itself. The shelf-
like surface illustrates how convergence can occur before reaching a minimum
energy conformation.
 Shallow minima are poorly defined, and present no obvious conformation on
which to converge:

Always examine the "log" file written during energy minimization. This file typically
indicates the energy and gradient at each iteration, and provides error messages
associated with convergence problems. Never assume that convergence has
occurred unless so indicated in the "log" file. Verify that the final gradient does not
exceed the cutoff value that you specified. The choice of minimization algorithm,
gradient cutoff, and the use of multiple starting conformations can help to ensure
convergence in most situations.
Some of the more commonly available minimization algorithms are
classified as

A) Which do not use derivatives

a) Simplex method
b) Sequential Univariate method

B) Which use derivatives of the energy with respect to coordinates

1. First Derivative
a) Steepest Descents Method
b) Conjugate Gradient Method

2. Second Derivative
c) Newton Raphson Method
Derivative/Nonderivative? – Provide shape of energy
surface/ calculating relative populations of the various
minimum energy structures.

Molecular Mechanics/Quantum Mechanics

Time/Steps Required

Downhill / Uphill

Input – Experimental/Theoretical
 A) Simplex method

Simplex is a geometrical figure with M + 1 interconnected vertices, where M is

the dimensionality of the energy function.

For a function of two variables the simplex is thus triangular in shape. A

tetrahedral simplex is used for a function of three variables and so for energy
function of 3N Cartesian Coordinates the simplex will have 3N + 1 vertices; if
internal coordinates are used then the simplex will have 3N-5 vertices.

Each vertex corresponds to specific set of coordinates for which energy can be
calculated.

The simplex algorithm locates a minimum by moving around on the potential

energy surface in a fashion that has been likened to the motion of amoeba.
Three basic kinds of move are possible.

A) Reflection of the vertex with the highest value

through the opposite face of the simplex, in
an attempt to generate a new point that has a
lower value.
If new point is lower in energy than any other
point in the simplex then a reflection and
expansion move may be applied. When a
valley floor is reached then a reflection move
will fail to produce a better point.

B) Under such circumstances the simplex

contracts along one dimension from the
highest point.

C) If this fails to reduce energy then a third type

move is possible in which simplex contracts along
in all directions, pulling around the lowest point.
For function x2 + 2 y2 the simplex is triangular in shape
Suppose initial simplex contains vertices located at point 1(9,9), 2(11,9), 3(9,11) .
The values of the function at these points are 243, 283, 323, resp. The vertex with
highest function value is at (9,11) so in first iteration this point is reflected through
the opposite face of the triangle, to generate a point with co-ordinates (11,7) and
has a function value 219. The highest vertex is now at (11,9). Which is reflected
through the opposite face of the simplex to give the point (9, 7) where the function
has the value of 179

Simplex vertex contains M+1 vertices to explore whole energy surface

B) Sequential Univariate method

For each co-ordinate, two

new structures are
generated by changing the
current co-ordinate
(i.e. xi +  xi and xi + 2 xi )

The minimum point in this quadratic function is determined and the co-ordinate
is then changed to the position of the minimum.
When the changes in all the co-ordinates are sufficiently small then the
minimum is deemed to have reached, otherwise a new iteration is performed
DERIVATIVE METHODS

Provide useful information about energy surface.

Used by most popular methods.
The direction of the first derivative of the energy (the gradient) indicates where
the minimum lies and the magnitude of the gradient indicates the steepness of
the local slope.
The energy of the system can be lowered by moving each atom in response to
the force acting on it: the force is equal to minus the gradient.
Second derivatives indicate the curvature of the function, information that can be
used to predict where the function will change direction (i.e. pass through a
minimum or some other stationary point)
Derivative Methods Line Search or Arbitrary Step Approach?

A) Steepest Descents method Iterative procedure/ analytical function?

Both the gradients and directions

are orthogonal
The steepest descent method works as follows for our trial function,

f (x, y) = (x2+2y2).

Differentiating the function gives df = 2xdx + 4ydy and so the gradient at any point
(x, y) equals 4y/2x.

The direction of the first move from to (9.0, 9.0) is (-18.0, -36.0) and the equation
of the line along which are search is performed is y= 2x-9. The minimum of the
function along with this line can be obtained using Lagrange multipliers and is at
(4.0,-1.0).

The direction of the next move is the vector (-8, 4) and the next line search is
performed along the line y = -0.5x + 1. The minimum point along this line is (2/3,
2/3) where the function has the value 4/3.

The third point found by the steepest descents method is at (0.296, - 0.074) where
the function has the value 0.099.

These moves are illustrated in figure.

B) Conjugate Gradient Method

The conjugate gradient method produces a set of directions which does not
show the oscillatory behavior of the steepest descent method in narrow
vallies. In the steepest descent method the gradient and the direction of
successive step are orthogonal. In conjugate gradients the gradients at each
point are orthogonal but the directions are conjugate.

A set of conjugate directions has

the property that for a quadratic
function of M variables, the
minimum will be reached in M
steps.
SUMMARY :

A) Simplex

1. Not a gradient minimization method.

2. Used mainly for very crude, high energy starting structures.
3. Expensive in terms of computer time due to large number of energ
calculations which are required. (less used for quantum mechanics)
4. Often used in combination with a minimization algorithm.: a few
steps of the simplex method are used to refine the initial structure
and then a more efficient method can take over it.

B) Sequential Univariate method

Requires few steps than simplex method

Steepest Descent

1. Follows the gradient of the energy function ("b") at each step.

This results in successive steps that are always mutually
perpendicular, which can lead to backtracking.
2. Works best when the gradient is large (far from a minimum).
3. Tends to have poor convergence because the gradient
becomes smaller as a minimum is approached.

Conjugate Gradients or Powell

1. Remembers the gradients calculated from previous steps to

help reduce backtracking.
2. Generally finds a minimum in fewer steps than Steepest
Descent.
3. May encounter problems when the initial conformation is far
from a minimum.
Newton-Raphson or BFGS

1. Predicts the location of a minimum, and heads in that direction

2. Calculates (N-R) or approximates (BFGS) the second derivatives in "A."

3. Storage of the "A" term can require substantial amounts of computer

memory.
4. May find a minimum in fewer steps than the gradient-only methods.

5. May encounter serious problems when the initial conformation is far

from a minimum.

The choice of minimization algorithm should be matched to the problem at hand.

Which minimization method should I use?

The storage and computational requirements

Relative speed with which the various parts of the calculation can be performed
Availability of analytical derivatives and the robustness of the method
Algorithm switching strategies are summarized in the following table
THANK YOU!!