Page 188 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Category Climate Change/Natural Disaster
Model Training and The GeoGPT models are built upon a diversified portfolio, includ-
Fine-Tuning ing Llama (USA), Qwen (China), Mixtral (Europe), and DeepSeek
R1 (China), offering users the flexibility to choose the model best
suited to their region and needs. These models undergo a series of
post-training processes, which include Continual Pre-training (CPT),
Supervised Fine-Tuning (SFT), and Direct Preference Optimization
(DPO). [3][4][5][6] The training process is carried out in three stages:
• Continual Pre-training (CPT): This stage utilizes a diverse set of
geoscience-related corpora to obtain a solid specialized model
for geoscience.
• Supervised Fine-tuning (SFT): This stage enhances the model’s
ability to follow geoscience-specific instructions by incorporating
QA pairs labelled by geoscientists, along with those generated
from the training corpus in CPT stage.

Human Preference Alignment: This stage uses the Direct Preference
Optimization (DPO) with preference data labelled by large language
models to align model's responses with human expectations and
preferences.

Testbeds or Pilot Deploy- GeoGPT platform: https:// geogpt .zero2x .org [8]
ments The open platform for GeoGPT, offering intelligent services such as
literature analysis, data extraction, map chat, domain-specific Q&A,
etc.

2 Use Case Description

2�1 Description

GeoGPT is an AI system powered by large language models (LLMs) with retrieval-augmented
generation (RAG) technology, applicable to the monitoring and protection of marine
ecosystems. The GeoGPT models are built upon a diversified portfolio, including Llama (USA),
Qwen (China), Mixtral (Europe), and DeepSeek R1 (China), offering users the flexibility to
choose the model best suited to their region and needs. [3][4][5][6]
One of its key applications is in deep-sea mineral exploration. By incorporating mineral
prospecting prediction models and risk assessment models, GeoGPT helps define target
exploration areas, assess the economic viability of extracting mineral resources, and provide
impact analyse to support exploration decision-making. This approach helps minimize
unnecessary mining activities, thereby reducing ecological damage. The system currently
achieves an accuracy rate of over 60%.

Another major application of GeoGPT is in marine biological disaster monitoring. The system
incorporates microalgae species classification algorithms to accurately identify harmful algae,
analyze algae distribution patterns, detect red tide signals, and predict the timing, location,
and scale of red tide outbreaks, significantly improving monitoring and early warning efficiency.
Furthermore, GeoGPT evaluates the impact of global warming on microalgae distribution,
shedding light on how climate change influences the behavior, distribution, and interactions
of algae in ecosystems. This capability provides critical scientific support for aquatic ecosystem

152