Item

Abstract

Recent advances in global geodetic mass balance and velocity assessments open new possibilities for global glacier model calibration and initialization. As the amount of data available for calibration increases in volume and complexity, how to best combine all these heterogeneous observations to initialize a dynamically consistent glacier evolution model?<p>In this contribution, we present the Open Global Glacier Data Assimilation Framework (AGILE), which iteratively adapts control variables to minimize a cost function penalizing mismatch to observations (i.e. an 'inversion' of the observations). AGILE uses automatic differentiation (AD) and the machine learning framework PyTorch to obtain the control variable sensitivities for efficient minimization. The flexible nature of AD allows the combination of temporally and spatially heterogeneous observational sources with a variety of control variables (e.g. glacier bed heights, mass-balance parameters, initial ice thickness, ...). Moreover, it makes it possible to switch parts of the modelling chain with differentiable counterparts.<p>We will demonstrate the capabilities of AGILE with idealized experiments using a re-implementation of an existing numerical model of glacier evolution (the Open Global Glacier Model, OGGM). We define bed height as well as an initial ice thickness distribution in the past as control variables, and&nbsp;constrain our inversion with synthetic observations orientated towards globally available data sets. Our approach removes the need for equilibrium assumptions or an 'apparent mass balance' specification, both required in other model-based bed-height estimates. Finally, we will discuss the added value of AGILE and upcoming challenges for a global operational application.