AI for Fusion & Complex Physical Systems
Supporting researchers, engineers, and interdisciplinary builders leveraging world modeling, neurosymbolic AI, and physical simulation to solve critical bottlenecks in next-generation nuclear fusion and complex real-world physics.
Technical Vision
Verifiable, high-assurance intelligence for real-world physics.
Some of the hardest technical challenges on earth.
Controlling high-beta plasma, mastering extreme thermodynamic regimes, and navigating non-linear physical systems represent some of the hardest technical challenges on earth. Traditional deep learning alone lacks the causal reasoning and strict safety guarantees required for real-world physical deployment, while classic physics models struggle to operate at the speed of real-time control.
GAIA Lab develops next-generation AI architectures that bridge this gap. By combining world models capable of internalizing complex environment dynamics with neurosymbolic AI that embeds explicit physical laws, logic, and safety constraints directly into neural networks, we aim to solve long-standing bottlenecks in nuclear fusion. Our goal is to enable verifiable, high-assurance intelligence capable of modeling, predicting, and controlling complex real-world physics—accelerating the path toward safe, abundant fusion energy.
Alisa34 · BMBF
≈ 2.1 M€Approved and funded by the BMBF over the course of three years. Project start 02/2025.
KIKO · BMBF
≈ 2 M€Approved and funded by the BMBF over the course of three years. Project start 03/2025.
Compute & energy
1,000,000× lessUp to a millionfold reduction in computational power/energy for training and operation.
Selected publications
08Across arXiv, Zenodo and conference venues, 2024–2026.
Focus areas
GAIA-enabled systemically safe AI.
01 / Active Inference
Model-based, real-time, explainable risk analysis based on Active Inference
- First-principles metric “fully loads” uncertainty and risk
- Stakeholders can see likely system trajectories, given current knowledge
- Modelers can target effort towards most impactful model improvements
02 / Governance
Cybernetic AI gatekeeper
- Algorithmically “takes over” agent behavior if risk exceeds a threshold
- Basis for transparent, trustworthy AI governance by the collective
- Demonstrated in an autonomous vehicle gatekeeping multi-agent study — see the arXiv preprint.
03 / Alignment
Responsive empirical value convergence
- Stakeholder-defined risk metrics and world models remove the black-box around AI thinking
- Feedback loops converge to stakeholder values, while adapting to the environment
04 / Strategic decision-making
The GAIA Tech Tree: a neurosymbolic AI framework for strategic technological decision-making
- A living sim of nuclear energy’s open problems, indexed to actual research
- Built with Universität Stuttgart and FuseNet
- Neurosymbolic decision-making across fusion and fission
How GAIA works
From a repository of world models to a bounded knowledge economy.
Five stages
-
01
The Repository
A Github for world models bootstrapped then growing over time
-
02
Protocol
Defining the rules for communication, updating and behaviour of world models
-
03
Decision Support
Extract information from the GAIA world models for important decisions in your organization
-
04
Convergence
Akin to Darwinian processes models compete between each other for top score wrt. quality, accuracy and transparency
-
05
Knowledge economy
The knowledge explosion ensues driven by increasing capabilities within a safe space bounded by GAIA
Projects
Domain-specific solutions, built with industry.
01
The GAIA Tech Tree
A neurosymbolic AI framework for strategic technological decision-making — a living sim of nuclear energy’s open problems, indexed to actual research.
Details
02
AI Safety from First Principles
Investigating Free-Energy based risk metrics with agentic AI.
Details
03
AI Safety for Nuclear Decommissioning
Modern AI tools based on SOTA machine learning and neurosymbolic AI, to understand regulatory mechanisms and automize review processes.
Alisa34
04
Collective Intelligence for Energy Transition
An AI-supported Collective Intelligence platform to support onboarding processes within organizations through machine learning components.
KIKO
News
Latest from the lab.
Funding
Large Scale AI funding approved: We are happy to announce two upcoming projects in AI Safety and Collective Intelligence within Germany’s unique energy transition program.
More info
Partnership
Partnership Announcement: We are looking forward to our cooperation with Lazy Dynamics and leveraging their toolkits for Bayesian Inference in collaborative projects together with the GAIA Lab.
Publication
We have published a study on systemically safe AI via gatekeeping estimating the risk in autonomous vehicles scenarios based on Active Inference.
arXiv.org
Affiliations & partners
-
Q.ANT
-
Q.ANT
-
Universität Stuttgart
-
FuseNet
-
Lazy Dynamics
-
BMBF
-
SPAR
Contact
We partner with industry on critical sectors.
Energy, finance, nature, cities, and manufacturing. Tell us what brought you here.