Title: Geometry Optimization
1Geometry Optimization
Computational Chemistry 5510 Spring 2006 Hai Lin
2Potential Energy Surface (PES)
A hypersurface in 3N-dimensional hyperspace,
where N is the number of atoms.
3Typical Tasks for Geometry Optimization
- Find a Local Minimum Structure
- Find the Global Minimum Structure
- Find the Transition State Structure
- Find the Reaction Pathway Connecting Two Minima
and Passing through the Transition State
Structure (will be talked about later this
semester)
4Gradients
- Gradient gx ?E/?x
- First-order derivatives of energy w/r variables
(e.g., Cartesian coordinates X, Y, Z, or
internal coordinates such as bondlength and angle
displacements) - The negative of the gradient is called force.
- Stationary Point
- A point on the PES where gradient is zero
- Including minimum, maximum, and saddle points
- (A transition state is a one-order saddle point.)
E
x
5Hessians
- Hessian Hxy ?2E / ?x?y
- A matrix of second-order derivatives of the
energy w/r variables (e.g., Cartesian or internal
coordinates) - Sometimes called force matrix
- Normal-mode Analysis
- Diagonalization of the Hessian matrix to obtain
vibrational frequencies (related to the
eigenvalues) and normal modes of vibrations
(eigenvectors) - Note Strictly speaking, normal-mode analysis is
only meaningful at stationary points.
6Minima, Maxima, Saddle Points
- A normal-mode analysis (also known as frequency
calculation) can be used to identify a stationary
point. - Minimum All frequencies are real (eigenvalues of
the Hessian are positive). - Maximum All frequencies are imaginary
(eigenvalues of the Hessian are negative). - Saddle point Some frequencies are imaginary.
- In particular, a saddle point with only one
imaginary frequency is called a transition state.
7Locate a Local Minimum
- Zero-order methods (energy only seldom used)
- First-order methods (energy and gradient)
- Second-order methods (energy, gradient, and
Hessian) - Use of more energetic information usually makes
it easier to find a minimum at a cost of higher
computational effort. - In numerical calculations, due to the finite
precision, a zero gradient is not exactly zero
and is smaller than a pre-defined cut-off value.
(Loose, tight, very tight ...)
8Steepest Descent Method
- A first-order method
- Take a step in the direction of negative gradient
- d - g
- The step size can be fixed or depend on the norm
of gradient.
- Once the energy increases, perform a line search
using the points in the last three steps to
determine where to go for the next step.
4
5
9Steepest Descent Method (2)
- Guarantee lowing the energy (when step size
dependeds on gradient) - Oscillating around the minimum energy path
- Never reach the minimum
- An ever descreasing speed approaching the minimum
- Relax a poor initial geometry quickly, and is
very useful for preliminary optimization
10Conjugate Gradient Methods
- Instead of follow the direction current gradient,
it follows a direction conjugate to previous
searching directions - di - gi bi di 1
- where bi is determined by gradients at the
current and the last or last few points. - Much better convergence characteristics than the
steepest descent method - Need to store the history Implicitly use
information about second-order derivatives - Used quite often, especially for biomolecules
11Newton-Raphson Methods
- Second-order methods
- Taylor expansion around a given point
- f(x) ? f(x0) gT(x x0) ½ (x x0)T H (x
x0) - ?
- (x x0) g / H - H1 g
- The displacements of the coordinates for the next
step follow the eigenvectors of the current
Hessian, and the step size following the i-th
eigenvector is determined as - Dxi fi / ei
- where fi is the projection of gradient along the
i-th eigenvector, whose eigenvalue is ei.
12Newton-Raphson Methods (2)
- Converge very efficiently, especially when close
to stationary points - Converge to the nearest stationary point,
regardless it is a minimum, maximum, or saddle
point. - A Trust Radius is used to prevent too large step
sizes that move the system out of the PES region
where the Tylaor expansion is valid.
- A shift parameter l can also be introduced to
modify the Hessian in order to control step
sizes - Dxi fi / (ei - l)
13Newton-Raphson Methods (3)
- Hessian can be calculated exactly, but usually
very expensively. Alternative solution is to use
an approximated Hessian that is updated (refined)
with energy and gradient informations on the fly
during the optimization. - The NR methods are not practical for optimizing
large-size systems, because calculating and
handling very large Hessians is difficult. (How
many Hessian elements do you have for a protein
of 10,000 atoms?)
14Restrained Optimization
- Keep the variable close to a desired value.
- Impose a penalty function, often a harmonic
potential function, to the energy expression - E E kp(x x0)2
- The force constant kp can be initially set very
large and descreases gradually. - Example
- Relaxation of a X-ray
- structure of a protein.
E
x0
x
15Constrained Optimization
- Satisfy certain constraints.
- G(x1, x2, ..., xN) 0
- Example keep the reaction coordinate defined by
a bondlength to a desired value (r r0 0). - Lagrange undetermined multipliers
- L(x1, x2, ..., xN, l) E(x1, x2, ..., xN) l
G(x1, x2, ..., xN) - where E(x1, x2, ..., xN) is the energy function
of the system. - Now we require
- ?L / ?xi 0, ?L / ?l 0
r
16Choice of Coordinates
- Cartesian Coordinates
- Straightforward, non-redundant, may not be
efficient due to high correlations, but sometimes
the only choice for large-size molecules - Internal Coordinates
- Close to chemical intuitions, usually redundant,
cause difficulty for certain geometries (e.g., an
angle accidentally close to 180 deg), need a
scheme to automatically generate for large
molecules - Mixed Cartesian and Internal Coordinates
- Convenient to use, not available in most programs
17Combinatorial Explosion Problem
The number of local minima typically goes
exponentially with the number of variables
(degrees of freedom).
Possible Conformations (3n) for linear alkanes
CH3(CH2)n1CH3
f
n 1 3 n 2 9 n 5 243 n 10 59,049 n
15 14,348,907 n 100 ?
f1
f2
18Locate the Global Minimum
- Systematic (Grid) Search (Manual or automatic)
- Monte Carlo (Stochastic) Methods
- Randam jumps on the PES
- Molecular dynamics
- Moving around based on Newtons laws of motion
- Simulated Annealing
- Cooling a melt slowly to form a crystal
- Genetic Algorithms
- Darwins principle survival of the fittest and
carry on to the next generation
19Monte Carlo Methods
20Monte Carlo Methods (2)
- Sample the PES only.
- Provide no time-dependent information.
- Special techniques are often used to promote a
reasonable acceptance ratio. - Energy minimization following distorted geometry
can be used to increase efficiency. - Geometry distortion is based on the last geometry
or on a set of selected previous geometries.
21Molecular Dynamics
- Sample the 6N-dimensional (kinetic potential)
hyperspace. - Provide time-dependent information.
- Special techniques can be used to promote a large
step size of time. - The total enegy set the limit of barrier heights
that can be overcomed. - Energy minimization is done for geometry
snapshots extracted from the MD trajectory.
22Simulated Annealing
- High temperature
- Overcome high barriers
- Sample large area
Gradually decease the temperature in the MC or MD
run
- Low temperature
- Overcome low barriers
- Sample small area
- If the annealing process is infinitely slow (not
possible in practice), the system reaches the
global minimum.
23Genetic Algorithms
Select a number of optimized conformations
(parents), calculate their fitness value
Define a fitness function f
Use the survived conformations as the next
generation (new parents)
Generate new conformations (children) by
crossover, mutation, ..., followed by
optimizations
Select conformations of highest fitness values
from parents and children
Calculate fitness values for children
Seems robust in finding minima pretty close to
the global minimum becomes popular in recent
years.
24Summary
- Typical Tasks for Geometry Optimizations
- Describe the PES
- Gradient, Hessian, Minima, Maxima, Saddle Points
- Find a local minimum
- Steepest Descent, Conjugated gradient,
Newton-Rasphon - Restrained and Constrained Optimizations
- Choice of Coordinates
- Find the global minimum
- Monte Carlo, Molecular Dynamics, Simulated
Annealing, Genetic Algorithms
25Your Homework
- Read Textbook (Remember to take notes!)
- 14.1, 14.2. 14.3
- 14.4
- 14.6
- 14.7.1, 14.7.2, 14.7.3, 14.7.4
26Practise with Tinker
- Optimization Programs Provided by Tinker
- For local minimum
- minimize (use energy and gradient)
- Newton (use energy, gradient, and Hessian)
- For global minimum
- Anneal (simulated annealing)
- Monte (Monte Carlo)
27Practise with Tinker (2)
- A small piece of peptide with known c5, c7a, and
c7e conformations. - (Coordinate files under the test subdirectory)
- Calculate the single-point energy for all three
geometries. Which one is of the lowest energy? - Optimize all three geometries using programs
minimize and Newton. Which one is of the lowest
energy? - Distort the geometries by editing manually the
coordinate files (.xyz), repeat step 2.
28Practise with Tinker (3)
- A small piece of peptide with known c5, c7a, and
c7e conformations. - Optimize all three geometries (c5, c7a, c7e)
using programs monte and anneal. Can you get the
global minimum? - Distort the geometries by editing manually the
coordinate files (.xyz), repeat step 4.
29Your Home work
- Try to find the global minimum or a minimum that
is close to the global minimum for - methane
- ethane
- butane
- octane
- long-chain n-alkane
- The files are named methane, ethane, butane,
octane, and alkane under the test subdirectory.
Use the MM3 force field.