Logistics AI with spatial intelligence

Artificial intelligence is already making the work of logistics managers and supply chain managers easier, for example with shipping notifications, returns analysis, picking sequences or freight cost development. All current language models are based exclusively on the interpretation of texts. Large Language Models (LLM) therefore search in vain for complete simulations of new warehouse locations or micro-hubs and their influence on delivery times and costs. They lack the space-time continuum, the geo-intelligence resulting from spatial information.

Changing geographies as a challenge

When data changes, for example a city grows, new construction sites are built or infrastructure expands, a standard LLM considers each update as new, independent information. A coherent landscape of relationships with changes over space and time remains unrecognized. Current systems work with text or object-based data and without a unified spatial framework. Even advanced AI systems without complex external GIS logic fail to track changing geographies, identify emerging hotspots or recognize spatial patterns over time. Neither Microsoft nor Google, OpenAI or Meta have spatio-temporal capabilities in their models. This requires a universal spatial language in addition to a large language model.

Machine-readable relationship network

To overcome this challenge, Prof. Dr. Roman Brylka and his team are constructing the Large Geo-Grid Model (LGM). The foundation is laid by the EEA Reference Grid, a Europe-wide standardized open reference grid of the European Environment Agency. TGB transfers it to the global space. In order to make spatial data comparable, superimposable and statistically analyzable, the model divides the earth into a precise, hierarchical grid of fixed cells. Each unit is given a permanent ID, a defined neighborhood and a consistent parent-child relationship across 17 grid levels with a scale of 1 mm to 100 km. As soon as the required data on depots, traffic situation, weather and fleet are projected onto this grid, each cell is understood as a consistent spatial token with a time stamp. Its characteristics evolve over time. The LGM learns the spatial grammar independently: How does each cell connect to its neighbors? How do layers aggregate from bottom to top? How are distances and orientations related? If data from different times is incorporated, the model recognizes spatial dynamics. It tracks growth fronts, identifies diffusion patterns and sees clustering trends. The LGM understands what is changing, where, how fast and in what direction.

The aim: logistics players promptly answer complex questions and receive intelligent and useful answers at lightning speed. At which points in the supply chain are there recurring structural delays and what alternatives are there? Which detours cost the most money and CO₂ each year? Which locations are suitable for temporary peak-season hubs?

A Large Intelligence Model (LIM) needs the LLM as a level of thinking and communication, while the LGM provides the spatial and temporal basis. In this configuration, the language model sends structured tool calls to the LGM via prompts. The LGM performs the geospatial inference and returns results in a form that the LLM can interpret and explain. The LIM (LLM + LGM) consults parameters such as infrastructure development, traffic data, weather conditions, routes and events, calculates densities, analyzes spatial patterns and provides an explanation and a map in response.