(491b) Optimizing Compiler For Parallelized Quantitative Kinetic Modeling Of Single-Site Olefin Polymerization

The quantitative modeling of kinetics is important for a number of significant chemical processes such as single-site olefin polymerization, vulcanization, oil refining, combustion and atmospheric chemistry, etc., where a large number of individual species can participate in a combinatorially large number of reactions. For example the description of single-site olefin polymerization requires keeping track of over 100,000 different species, resulting in extremely large sets of ODEs which contain of order ten different rate constants that must be regressed from experimental data. The real challenge is that the exact kinetic mechanism is not know a priori for a specific catalytic system and must be determined. The researcher must discriminate between several different mechanisms for each kinetic step and since the overall polymerization model has a number of kinetic steps, there are potentially several thousand different models. The creation of just a single model of this complexity is not easy, but the generation, management and parameter optimization of thousands of these models is impossible with out significant cyber-infrastructure.

To enable the modeling of experimental batch polymerization data, a suite of cyber-tools has been developed so that the researcher can formulate a new model in a page of near English language rules, where the resulting large ODE set is automatically generated and then employed in a parallel optimization code for model discrimination and parameter estimation. The system includes: (i) a Chemical Compiler that allows the researcher to specify chemical reactions in a near English language and generates the complex set of reactions; (ii) a Model Selection Processor, where the specific model of interest is selected by the researcher; (iii) an Equation Generator which efficiently generates 1000's of ODEs; (iv) an Equation Optimizer that uses algebraic laws and common sub-expression elimination to reduce the computational burden; and, (iv) a Parallel Optimization Module that uses a parallelized Levenberg-Marquardt algorithm, takes advantage of the data structure and uses dynamic load balancing for 90% efficiency on parallel machines. Using these tools, our group can now formulate a new kinetic mechanism, automatically generate the complex computer code and fit the model to experimental data in several hours ? a process that previously took several months. This cyber-enable modeling capability allows us to analyze different model hypothesis and design new experiments to discriminate between candidate models at a rate that keeps pace with the rate of generation of experimental data.