Our research focuses upon the creation of adaptive software to solve computational
problems that are considered intractable to compute by direct techniques.
Adaptive programs emulate the process of evolution by breeding populations
of solutions within a defined problem space, then retaining and combining
the individual solutions which best fit the problem space. The fields to
which these techniques have been applied to include :
- Electromagnetic fields and electronic systems
Generating and protecting against the effects of picosecond pulsed electromagnetic
fields is of particular relevance to satellite and military electronic
and radar stealth systems. We have designed programs to evolve structures
for the generation and propagation of broadband electromagnetic pulsed
fields. This has resulted in the discovery of some unexpected configurations,
including novel types of spark switched Hertzian devices and spiral
TEM antennas. Using similar techniques to evolve shielding we are finding
novel means to protect electronic devices from powerful electromagnetic
fields. These include conductive fibers suspended in dielectrics with
carefully selected properties.
These programs use time domain solutions to the Maxwell equations and
evolve topologies from an initial population of random dielectric/conductor
configurations. In one case the goal is the to generate the highest
directional electromagnetic pulse intensity for a given amount of input
energy in the shortest time, in the other the goal is to dissipate or
absorb without reflecting the most electromagnetic energy impinging
upon the structure from a baseband source.
- Predicting the physical properties of materials
If it were possible to accurately predict the physical properties
of a molecule we could screen any molecule for desired properties using
computers. It would then be possible to create synthetic medicines,
electronic components, polymers, alloys etc. within a computer and then
test them for required properties. In principle it is possible to predict
every property of any molecule or atom by finding the exact solution
to the Schrödinger equation for the electron wave of that molecule or
atom. However there is a severe and intractable problem relating to
solving this form of the Schrödinger equation.It simply cannot be solved
exactly for any system containing more than two interacting particles.
This means for any atom or collection of atoms other than monatomic
hydrogen we cannot with certainty predict the properties it will exhibit.
It is only possible to find an approximate a solution, and molecules
do not posses 'approximate' properties.
We are researching a promising alternative approach to predicting physical
properties. This approach is based upon finding the function that maps
directly from the structure of the molecule/atom to one of its properties
using adaptive software, then applying that function to new configurations.
The software evolves within a problem space consisting of training and
validation databases consisting of the physical properties and connected
graph descriptions of many thousands of molecules. The goal is to find
those functions which map from the graph of the molecule to the physical
property with the smallest error over the both the training and the
validation databases. This process is then repeated until the population
converges on an accurate solution.
|