Experiences in High Performance Computing Used to Characterize Maize Phenology.

Novel proximal sensing technologies can rapidly characterize phenotypes expressed by crop genotypes in the field. While raw sensor outputs or simple indices calculated from such data can be analyzed as phenotypes, crop simulation modeling can potentially estimate phenotypes for more fundamental biological traits, which should exhibit reduced effects of environment and thus higher heritabilities. By using proximal sensing data to invert crop models, a transformative increase in the quality of phenotypic data could be realized. A key challenge is to identify practical methodologies for model inversion.  The objective of this work was to investigate the use of high-performance computing for this purpose.  The 'Stampede' supercomputer at the Texas Advanced Computing Center (TACC) was used to conduct approximately 350M simulations of maize development using CSM-CERES-Maize.  Required environmental and management data were obtained for 11 environments, which were collected in connection with field tests of the maize Nested Associated Mapping (NAM) population.  Anthesis dates for the maize NAM population, which includes over 5000 maize lines, were also obtained for each environment.  Five CSM-CERES-Maize parameters (P1, P2, PHINT, P2O, and TBASE) were systematically varied using a Sobol sampling scheme.  Parameter sets that minimized root mean squared error (RMSE) between measured and simulated anthesis date were identified for each maize line. The simulation exercise required 63,372 CPU-hours on the TACC computer and averaged 176 runs s^-1.  Equifinality issues led to non-uniqueness of optimum parameter sets (i.e., the optimum RMSE was identical for several parameter sets).  Current efforts focus on understanding the causes for parameter non-uniqueness and identifying the degree to which modeling artifacts hamper the extraction of biological meaning from the simulation results.