An autonomous phase diagram database for geodynamic simulations of magmatic systems.

Example of clustering algorithms applied to different phase diagram calculation.


Self-consistent modelling of magmatic systems is challenging as the melt continuously changes its chemical composition upon crystallizing, which may affect the mechanical behaviour of the system. Melt extraction and subsequent crystallization creates new rocks while deplet- ing the source region. As the chemistry of the source rocks changes locally due to melt extraction, new calculations of the stable phase assemblages are required to track the rock evolution and the accompanied change in density. As a consequence, a large number of iso- chemical sections of stable phase assemblages are required to study the evolution of magmatic systems in detail. As state of the art melting diagrams may depend on 9 oxides as well as pressure and temperature, this is a 10-dimensional computational problem. Since computing a single isochemical section (as a function of pressure and temperature) may take several hours, computing new sections of stable phase assemblages during an ongoing geodynamic simulation is currently computationally intractable. One strategy to avoid this problem is to precompute these stable phase assemblages and to create a comprehensive database as a hyperdimensional phase diagram, which contains all bulk compositions that may emerge during petro-thermo-mechanical simulations. Establishing such a database would require repeating geodynamic simulations many times while collecting all requested compositions that may occur during a typical simulation and continuously updating the database until no additional compositions are required. Here, we describe an alternative method that is better suited for implementation on large scale parallel computers. Our method uses the entries of an existing preliminary database to estimate future required chemical compositions. Bulk compositions are determined within boundaries that are defined manually or through princi- pal component analysis (PCA) in a parameter space consisting of clustered database entries. We have implemented both methods within a massively parallel computational framework while utilizing the Gibbs free energy minimization program Perple X. Results show that our autonomous approach increases well the resolution of the thermodynamic database in com- positional regions that are most likely required for geodynamic models of magmatic systems.

Geophysical Journal International