Alisa34 | Project Outline

With regulations in place covering every legal aspect of any process in the nuclear industry, the research project Alisa34 aims at easing the process of nuclear decommissioning - one of a kind in terms of scale and complexity. The expectation level towards personnel couldn't be higher: Rules and regulations about (de-)commissioning of nuclear facilities date back decades into the 1980s and even further with knowledgable workforce retiring, which, together with a sustained talent gap, makes this mix a massive challenge for Germany's ambition in regards to its energy transition.

Alisa34 aims to tackle this problem with the help of AI and its potential to collect, memorize and replicate patterns found in large text corpora. In order to solve the challenge and improve application processes within nuclear decommissioning by (partial) automization, Alisa34 will be based on four components:

1. Data collection and de-siloing

2. Large Language Models for information retrieval

3. AI Safety components for machine-learning whiteboxing

4. Agent-based system for process automation and stakeholder management

The foundation of the project is built on top of a large document set of partially structured data. Many documents are properly structured, however typically spread across many filesystems and different versions, moreover potentially interpreted and updated by various stakeholders over the course of time (thus potentially inducing personal bias). Integrating these into a clean data set to subsequently create ontologies and properly structured rule sets will be laying the groundwork.

Subsequently, the next goal is to build a symbolic, testable world model reflecting the logic of the application process and other relevant domain knowledge such as nuclear physics and civil engineering. This model will 'understand' the logic of an application process and be integrated into stakeholders’ workflows in order to flag conundrums as well as find redundancies or similarities, so as to either avoid logical fallacies or, even better, speed up processes. To this end, we will develop a neurosymbolic framework, using fine-tuned Large Language Models for automated data processing, knowledge extraction into formal/causal models, and automated document creation, as well as active inference metamodels to find solutions in the world model, identify data and knowledge gaps, and optimize applications.

This framework naturally integrates AI safety concerns, ensuring robust AI behavior by making the decisions of AI modules transparent, verifiable and auditable.

In order to inform stakeholders about relevant changes in the process, applications/documents or regulations, agent-based AI modules will ensure that the right people are connected and informed about relevant changes, and subsequently processes can already be started and even partially completed.

In the final stage, Alisa34 will deliver a fully-fledged machine-learning aided work environment with neurosymbolic AI taping into every reelvant process step (cf. Image 2). The human stakeholder shifts from manual labor into an oversight role with AI taking over the bulk of fact-checking and knowledge acquisition + verification steps. Documents can be created in a fully automated fashion with agent-based systems shipping relevant knowledge bits in between stakeholders to inform, verify and plan process steps.