2006 MCS Divisional Seminars & Colloquia

Active Thermochemical Tables: A Vision for Petaflop Thermochemistry

   Branko Ruscic

 Chemistry Division, Argonne

  Hosted by  Dmitry Karpeev

3:00 PM, February 1, 2006
Building 221,  Room A216

The general concept of Active Tables (AT) will be introduced and elaborated using Active Thermochemical Tables (ATcT) as a concrete example. ATcT are a new paradigm that catapults thermochemistry from the traditional sequential approach into the digital computing arena of the 21st century. The underlying Thermochemical Network (TN) approach, which exposes and fully utilizes the maze of inherent multiple interdependencies that is largely ignored or simplified by the traditional approach with a resulting loss of knowledge content, allows ATcT to provide reliable, accurate, and internally consistent thermochemistry needed in many domains of chemistry. The ATcT approach involves, inter alia, a statistical evaluation of experimental and theoretical determinations that define the TN. This is made possible by redundancies in the TN, such as competing measurements and alternate network pathways. The statistical analysis produces a self-consistent TN, from which the optimal thermochemical values are obtained by simultaneous solution in error-weighted space, thus allowing the best possible use of all the knowledge present in the TN.

Besides deriving superior results by fully utilizing the available knowledge, ATcT offer a number of additional features that are neither present nor possible in the traditional approach, such as full propagation of new knowledge to all affected results, hypothesis testing and evaluation, discovery of weak links in the TN that provide pointers to new experimental or theoretical determinations, etc. The constantly growing TN has evolved to the current a set of ~750 species interlinked by ~7,000 determinations, relevant to combustion processes. Its solution already involves unacceptably long execution times on commodity computers, posing an imminent need to transition to Teraflop resources. However, based on current timing measurements, even teraflop resources will become inadequate once the TN reaches ~2500 species.

Projections show that our goal TN, which would contain ~10,000 species and would come closer to satisfying the more general thermochemical needs in chemistry, would have execution times for a single run of the order of a millennium on a commodity computer, and almost two years on a teraflop machine, but it would take less than a day on a petaflop machine. While there is room for both improving the procedure and re-optimizing various portions of the current code, which may result in a non-negligible linear factor of speedup, the basic bottleneck is in the underlying ~n⁴ scaling of the problem, which quickly outruns any linear improvements.