GeoTransformer

Status: Running

A research project to revolutionize geostatistical simulation with transformer architectures.

geostatistics machine learning transformers stochastic simulation spatial modeling AI

The GeoTransformer project aims to revolutionize geostatistical simulations by integrating artificial intelligence transformer models — a technology that has significantly advanced natural language processing.

Traditional geostatistical methods (e.g., kriging, multiple point geostatistics) often struggle with complex spatial structures, non-stationarity, and limited training data. GeoTransformer addresses these limitations through transformer architectures adapted for spatial data. Key innovations include:

  • Spatial encoding techniques tailored for multidimensional data
  • Continuous variable generation to support realistic simulation of physical phenomena
  • Retrieval-augmented generation (RAG) to incorporate new training images on the fly
  • Multi-variable simulations via cross-attention mechanisms

This project will improve the scalability and realism of simulations, with applications in:

  • Climate science
  • Natural hazard assessment
  • Precision agriculture
  • Land-use planning

Conducted in collaboration with Grégoire Mariethoz (University of Lausanne), GeoTransformer bridges AI and geostatistics to deliver high-fidelity, adaptive simulation tools for scientific and operational use.

Team Members