Upcoming Event: Jackson School of Geosciences & Oden Institute

Monster Codes or Modular Differentiable Software in Computational Geosciences? Applications to Magmatic Systems and Industry

Boris Kaus, Johannes Gutenberg Universität Mainz

Abstract

Understanding the dynamics of geoscientific processes requires numerical models that account for the underlying physics of the involved processes. Over the past few decades, large, parallel, community codes have driven significant advancements in computational geosciences. However, these monolithic codes, often written in low-level languages, optimized for CPU architectures, and comprising hundreds of thousands of lines of code, are difficult to modify and maintain. This hinders innovation, as much ongoing and future research requires incorporating additional physics or linking simulations with observations (for example, through adjoint-based inversion approaches).  

Magmatic and volcanic processes provide a key example of a complex system; they operate across a wide range of length and timescales, and experience phase transitions and changes in chemistry. Despite increasing capabilities for monitoring volcanoes (such as tracking ground deformation and earthquakes), linking observations to the physics of magmatic systems remains challenging, as no integrated computational model of magmatic systems currently exists. While certain parts of the problem are well understood through dedicated models and experiments, a comprehensive magmatic system model is needed to assess the relative impact of individual processes on the larger system. Importantly, the computational framework must be flexible enough to easily integrate new processes or machine learning approaches. 

Here, I will give an overview of our recent work in developing modular software packages in the Julia language as a replacement for monolithic community codes. These modular packages offer significant advantages: they are easier to maintain and extend, work seamlessly on GPUs, and accommodate nonlinear constitutive relationships regardless of the underlying solver discretization (finite element vs. finite difference). We are using this to develop magmatic system models that run on parallel GPU clusters and account for the visco-elasto-plastic deformation of rocks. 

To simulate the drastic changes in magma chemistry and viscosity during crystallization, we have developed an efficient computational thermodynamics code that integrates seamlessly with the thermo-mechanical models. Additionally, to assess the impact of model uncertainties on (surface) observations, we employ adjoint-based sensitivity kernels, computed using automatic differentiation tools. This enables a routine assessment of the impact of model uncertainties, even when the underlying model physics changes significantly.

To demonstrate the broad applicability of these computational tools, I will also present examples of their use in problems relevant to industry, such as determining the long-term stability of salt caverns and creating digital rock physics models.

 

Biography

As a geodynamicist Boris' main interest is understanding how geological processes work, from the grain scale to the scale of a planet. This is predominantly done with the help of mathematical and numerical models, for which we develop new approaches in our group. Ongoing research projects include lithospheric-scale shear localization, the coupling between melt migration and lithosphere deformation, the formation of fold-and-thrust belts, the coupling between erosion, lithosphere dynamics and mantle flow, the dynamics of subduction zones as well as the development of new software that runs on high-performance computing systems.

His teaching interests include geophysics, modelling of tectonic processes, the dynamics of the earth and lithosphere as well as anything related to quantitative geosciences.