Transcript Document

The Materials Computation Center, University of Illinois
Duane Johnson and Richard Martin (PIs), NSF DMR-03-25939 • www.mcc.uiuc.edu
Accurate Semi-Empirical Quantum Chemistry via Evolutionary Algorithms
Duane Johnson, Todd Martinez, and David Goldberg and
OBJECTIVE: To accelerate Quantum Chemistry
(QC) simulations of excited-state reactions by
+1000 times by creating semi-empirical
potentials with accuracy approaching that of
high-level methods.
APPROACH: Use machine-learning methods based
upon efficient, Competent Genetic Algorithms (eCGA)
and multi-objective optimization (MO).
WHY IT MATTERS: With fast but accurate semiempirical potentials, we can search for new drugs or
critical biological reactions 100 – 1000 times faster!
STRATEGY: Using well-known semiempirical (MP3)
QC potential* we optimize two objectives (error in
energy and energy-gradient) for ethylene (C2H4)
using a few excited-state structures calculated from
high-level QC (ab initio CASSCF learning set) to
predict excited-states not in the learning set.
*MP3 potential has 11 parameters just for carbon.
• We get agreement with high-level QC methods:
160±40 fs (vs.180±50 fs) for excitation decay.
D2d excitation energy 2.5 eV (vs 2.5 eV).
Pyramidalization energy 0.9 eV (vs 0.9 eV).
Potentials 1000x faster and “transferable” to
other C-H based molecules.
Students: Kumara Sastry and Alexis Thompson
E
Pyramidalized
D2d twisted
o Stable Solutions
+ Sensitive/Unstable
Physically
most optimal
RESULT: (upper) An excitation of ethylene
and example excited-state configurations.
(lower) Multi-objective Genetic Algorithms
found the “best MP3 solutions” (circles) with
“on-the-fly” sensitivity analysis of solution sets.
 All optimal solution sets agree with our highlevel QC calculations.
The Materials Computation Center, University of Illinois
Duane Johnson and Richard Martin (PIs), NSF DMR-03-25939 • www.mcc.uiuc.edu
OUTLOOK: We are finalizing analysis for ethylene and
benzene and details of why “non-dominate” Pareto
front and eCGA are necessary to do well, as opposed
to standard GA’s being used in chemistry.
We have:
• provided MO-GA code on Software Archive©.
• proven the utility of MO-GA using eCGA.
• shown transferability of potentials for other C-H
molecules that were not used in learning set.
• verified the cusp surfaces of the excited molecules
are described well by semi-empirical potentials.
• revealed the importance of “non-domininant Pareto
fronts”, “crowding distances”, and “tournament
selection” to obtain good MO-GA solutions.
PUBLICATIONS 2006-2007:
• Kumara Sastry, D.D. Johnson, Alexis L. Thompson,
D.E. Goldberg, T.J. Martinez, "Optimization of
Semiempirical Quantum Chemistry Methods via
Multiobjective Genetic Algorithms: Accurate
Photochemistry for Larger Molecules and Longer
Time Scales" (invited) Materials and Manufacturing
Processes 22 (2007) 553 - 561.
• Kumara. Sastry, D.D. Johnson, and D.E. Goldberg,
"Scalability of a Hybrid Extended Compact Genetic
Algorithm for Ground State Optimization of
Clusters,” (invited) Materials and Manufacturing
Processes 22 (2007) 570 - 576.
• Kumara Sastry, D.D. Johnson, Alexis L. Thompson,
D.E. Goldberg, T.J. Martinez, J. Leiding, and Jane
Owens, "Multiobjective Genetic Algorithms for
Multiscaling Excited-State Dynamics in
Photochemistry," GECCO 1745-1752 (2006)
*Silver Medal, Best Paper in real-world track.
Recognition and Industry:
• At Genetic and Evolutionary Computation Conf. 2006.
 Awarded Silver “Hummie” Medal.
 Awarded “Best Paper” in Real-World Applications.
• Student Kumara Sastry was finalist for the LemelsonMIT innovation prize.
• US provisional patent application made.
 K. Sastry joining INTEL in fall 2007 to develop futuregeneration chip via optimization.