How AI is Powering Up the Nuclear Industry 

Sequoyah Nuclear Power Plant 

In an era where digital fluency is the new literacy, Large Language Models (LLMs) have emerged as revolutionary game-changers. These models are not just regurgitating information; they’re learning procedures and grasping formal logic. This isn’t an incremental change; it’s a leap. They’re making themselves indispensable across sectors as diverse as finance, healthcare, and cybersecurity. And now, they’re lighting up a path forward in another high-stakes arena: the nuclear sector.

The Limits of One-Size-Fits-All: Why Specialized Domains Need More Than Standard LLMs

In today’s digital age, Large Language Models (LLMs) like GPT-4 have become as common as smartphones, serving as general-purpose tools across various sectors. While their wide-ranging training data, which spans from social media to scientific papers, is useful for general capabilities, this limits their effectiveness in specialized domains. This limitation is especially glaring in fields that require precise and deep knowledge, such as nuclear physics or complex legal systems. It’s akin to using a Swiss Army knife when what you really need is a surgeon’s scalpel.

In contrast, specialized fields like nuclear engineering demand custom-tailored AI solutions. Publicly-available LLMs lack the precision needed to handle the nuanced language, complex protocols, and critical safety standards inherent in these areas. Custom-built AI tools go beyond mere language comprehension; they become repositories of essential field-specific knowledge, imbued with the necessary legal norms, safety protocols, and operational parameters. By focusing on specialized AI, we pave the way for more reliable and precise tools, moving beyond the “Swiss Army knife” approach to meet the unique demands of specialized sectors.

LLMs are Swiss Army knives in that they are great at a multitude of tasks; this is paradoxical to their utility in a field like nuclear where nuance is everything.

The Swiss Army Knife In Action

Below is a common response from a public chatbot on most plant specific questions. The information about this site is widely available online and has been published well before 2022 with the power plant’s commission date occurring in 1986.

From the chatbot’s response, the generic information provided by this public-available model does not give enough clarity for experts to rely on. To answer the above question, the model will need to be adapted to a specific domain.

Adapting general models to be domain specific is not easy however.  Some challenges with this task include:

  1. Financial and Technical Hurdles in Fine-Tuning—Fine-tuning public models is a costly affair. Beyond the financial aspect, modifications risk destabilizing the intricate instruct/RLHF tuning, a nuanced balance established by experts.
  2. Data Security: A Custodian Crisis —Public models weren’t built with high-security data custodianship in mind. This lack of a secure foundation poses risks, especially for sensitive information.
  3. A Dead End for Customization—Users face a brick wall when it comes to customizing these off-the-shelf models. Essential access to model weights is restricted, stifling adaptability and innovation.
  4. Stagnation in Technological Advancement —These models lag behind, missing out on revolutionary AI developments like RLAIF, DPO, or soft prompting. This stagnation limits their applicability and efficiency in evolving landscapes.
  5. The Impossibility of Refinement and Adaptation—Processes integral for optimization, such as model pruning, knowledge distillation, or weight sharing, are off the table. Without these, the models remain cumbersome and incompatible with consumer-grade hardware.


NuclearN specializes in AI-driven solutions tailored for the nuclear industry, combining advanced hardware, expert teams, and a rich data repository of nuclear information to create Large Language Models (LLMs) that excel in both complexity and precision. Unlike generic LLMs, ours are fine-tuned with nuclear-specific data, allowing us to automate a range of tasks from information retrieval to analytics with unparalleled accuracy.

What makes our models better than off-the-shelf LLMs? 

Large Language Models (LLMs) from NuclearN are trained on specialized nuclear data that are transforming several core tasks within the nuclear industry, leveraging their vast knowledge base and advanced understanding of nuclear context-specific processes. These models, when expertly trained with the right blend of data, algorithms, and parameters, can facilitate a range of complex tasks and information management functions with remarkable efficiency and precision.

NuclearN is training our LLMs to enhance several core functions:

  1. Routine Question-Answering: NuclearN’s trains LLMs on a rich dataset of nuclear terminologies, protocols, and safety procedures. They offer accurate and context-aware answers to technical and procedural questions, serving as a reliable resource that reduces the time needed for research and minimizes human error.
  2. Task-Specific and Site-Specific Fine Tuning: Even though our LLMs are trained to be nuclear-specific, different sites can have very specific plant designs, processes, and terminology.  Tasks such as engineering evaluations or work instruction authoring may be performed in a style unique to the site.  NuclearN offers private and secure, site and task-specific fine tuning of our LLMs to meet these needs and deliver unparalleled performance.
  3. Neural Search: The search capabilities of our LLMs go beyond mere keyword matching. They understand the semantic and contextual relationships between different terminologies and concepts in nuclear science. This advanced capability is critical when one needs to sift through large volumes of varied documents—be it scientific papers, historical logs, or regulatory guidelines—to extract the most pertinent information. It enhances both the efficiency and depth of tasks like literature review and risk assessment.
  4. Document Summarization: In an industry awash with voluminous reports and papers, the ability to quickly assimilate information is vital. Our LLMs can parse through these lengthy documents and distill them into concise yet comprehensive summaries. They preserve key findings, conclusions, and insights, making it easier for professionals to stay informed without being overwhelmed by data.
  5. Trend Analysis from Time-Series Data: The nuclear industry often relies on process and operational data gathered from sensors in the plant to track equipment performance and impacts from various activities. NuclearN is training our LLMs to be capable of analyzing these time-series data sets to discern patterns, correlations, or trends over time. This allows our LLMs to have a significantly more comprehensive view of the plant, which is particularly valuable for monitoring equipment health and predicting operational impacts.

By leveraging the capabilities of NuclearN’s specialized LLMs in these functional areas, the nuclear industry can realize measurable improvements in operational efficiency and strategic decision-making.

Stay informed and engaged with everything AI in the nuclear sector by visiting The NuclearN Blog. Join the conversation and be part of the journey as we explore the future of AI in nuclear technology together. 

Nuclearn v1.8 – Neural Search and Easier Automation

Nuclearn recently released version 1.8 of its analytics and automation platform, bringing major upgrades like neural search for intuitive queries, configurable automation routines, expanded analytics outputs, and enhanced ETL data integration. Together these features, some of them AI-driven, aim to optimize workflows and performance.

Neural Search

The neural search upgrade allows searching based on intent rather than keywords, even with ambiguous queries. Neural algorithms understand semantics, context, synonyms, and data formats. This saves time compared to traditional keyword searches, and provides significant advantages when context-sensitive information retrieval is crucial.

Some of the benefits of neural search include:
Precision of Search Results: Traditional keyword-based searches often yield an overwhelming number of irrelevant results, making it difficult for plant personnel to find the specific information they need. Neural search engines deliver results with ranked relevance. This means results are not just based on keyword match but on the basis of how closely the content of the document matches the intent of the search query.  

Contextual Understanding: Boolean queries, which are typically used in traditional search engines, lack the ability to understand the contextual nuances of complex technical language often found in engineering and compliance documentation. Neural search algorithms have a kind of “semantic understanding” that can understand the context behind a query, providing more relevant results. In addition, Neural search understands synonyms and related terms, crucial when dealing with the specialized lexicon in nuclear, thus making searches more robust.

Multiple Data Formats: Nuclear plants often store data in different formats, such as PDFs, Word documents, sensor logs, and older, legacy systems. A neural search engine can be trained to understand and index different types of data, providing a unified search experience across multiple data formats. 

Selective Classification for Unmatched Automation Accuracy

AutoCAP Screener also saw major improvements in v1.8. You can now set desired overall accuracy levels for automation templates. The Nuclearn platform then controls the confidence thresholds using a statistical technique called “selective classification” that enables theoretically guaranteed risk controls. This enables the system to ensure it operates above a user-defined automation accuracy level.


With selective classification, plants can improve automation rates and efficiency without compromising the quality of critical decisions. Risk is minimized by abstaining from acting in uncertain cases. The outcome is automation that consistently aligns with nuclear-grade precision and trustworthiness. By giving you accuracy configuration control, we ensure our AI technology conforms to your reliability needs. 

Additionally, multiple quality of life enhancements were added to the AutoCAP audit pages. Users can now sort the audit page results, add filters, integrate PowerBI dashboards with audit results, and even export the automation results to csv. These enhancements make it easier and more flexible for users to assess, evaluate, and monitor the automation system.

Analytics & Reporting Enhancements

On the analytics front, our customers wanted more customizations. v1.8 answers their request with the ability to upload their own custom report templates. In addition, customers can change date aggregations in reports to tailor the visualizations for specific audiences and uses. Enhanced dataset filtering and exporting also allows sending analyzed data to PowerBI or Excel for further manipulation or presentation.


Editing analytics buckets is now more flexible too, with overwrite and save-as options. We added the ability to exclude and filter buckets from the visualization more easily and make changes to existing buckets, including their name.  

Data Integration

Behind the scenes, ETL workflows (meaning “extract, transform, load” data) were upgraded to more seamlessly ingest plant data into the Nuclearn platform. Users can now schedule recurring ETL jobs and share workflows between sites. With smooth data onboarding, you can focus your time on analytics and automation rather than manually uploading data. 

With advanced search, configurable automation, expanded analytics, and optimized data integration in v1.8, the Nuclearn Platform is better equipped to drive operational optimization using AI-powered technology. This release highlights Nuclearn’s commitment to meaningful innovation that solves real-world needs.

DARSA: The Guide to Full Process Automation Using AI

You don’t automate right away…

Process automation using Artificial Intelligence is a complex endeavor. To successfully automate a process, automation systems and their implementers need to effectively incorporate complex technologies, a deep understanding of the business processes, risk-based decision making, and organizational change management all at once! This challenge can feel insurmountable to many organizations looking to start adopting AI-driven process automation. And unfortunately for some, it has proven to be so.

Luckily, there are battle-tested methods for bringing an automation system to life and avoiding the potential pitfalls. Here at Nuclearn, we have developed a project implementation process we call DARSA that helps guide us through automation projects. DARSA helps us deliver maximum value with minimal risk by leveraging an iterative, agile approach to AI-driven automation.

So what is DARSA? DARSA is a five-step linear process that stands for Decisions-Data-Direction, Assess, Recommend, Semi-Automate, and Automate. Each step in DARSA is a distinct phase with distinct characteristics, and transitions between these phases are planned explicitly and usually require system changes. To learn more about DARSA, we must dive into the first phase: “Decisions, Data and Direction”.

1) Decisions – Data – Direction

Before starting AI-driven automation, it is important to specify several key factors that will guide the project. These items are Decisions, Data, and Direction.

Decisions are the first, and most critical item to define early in an AI-driven automation project. At the end of the day, if you are embarking on an AI-driven automation project, you are doing so because you need to automate challenging decisions that currently require a human. If there is no decision to be made, then AI will not help automation in any meaningful way. If the decisions are trivial or based on simple rules, there is no need for AI. These processes should be automated by traditional software. So the first, essential part of an AI-driven automation project is identifying and defining the decisions that are planned to be automated.

Data is the next item that must be specified. An AI-driven automation project needs to identify two key sets of data: the input data being used to influence decisions, and the data created as a result of those decisions. These are critical, as AI-driven automation relies heavily on machine learning. To learn how to make decisions automatically, machine learning requires historical data. For each decision, there should be a detailed log of all data used at decision time, and a log of all historical human decisions.

Direction. AI projects must begin with clear direction, but unlike large traditional software projects, AI-driven automation projects cannot begin with a detailed project plan laying out detailed requirements gathering and design. AI and machine learning are notoriously unpredictable – even the most experienced practitioners have challenges predicting how well models will perform on new datasets and new challenges. Automation systems often have to evolve around the unpredictable strengths and weaknesses of the AI. As a result, it is important to specify a clear direction for the project. All members of the project should be aligned with this direction, and use it to guide their iterations and decisions. For example, in an AI-driven automation system for helping generate Pre-Job Brief forms, the direction for the project might be “Reduce the amount of time required to assemble Pre-Job Briefs while maintaining or improving the quality of Pre-Job Briefs”. This simple statement of direction goes a long way towards bounding the project scope, ruling out system design decisions that are unacceptable, and fostering potential innovation.

2) Assess

Once the Decisions, Data and Direction are specified, the most important factors for success in an AI-Driven automation project are fast iterations and good feedback. That is why the second phase in DARSA is “Assess”. During this stage of the project, nothing is actually automated! The “automation” results are being shared with subject matter experts, so they can assess the results and provide feedback. Automation system designers and Data Scientists are already quite familiar with testing how well their system works via various traditional methods. While these can help with generating accuracy metrics (the model is 95% accurate!), these methods are quite poor at evaluating exactly where the automations will fail or be inaccurate, why they are that way, and what the impacts are. The Assess phase is often where important risks and caveats are identified, and where additional considerations for the project are discovered.

Let’s take for example my experience with a project attempting to automate the screening of Condition Reports (CR) at a Nuclear Power Plant. One of the key decisions in screening a CR is determining whether the documented issue has the potential to affect the safe operation of the plant, often referred to as a “Condition Adverse to Quality”. Before even showing our AI model to users, my team had produced some highly accurate models, north of 95% accurate! We knew at the time that the human benchmark for screening was 98%, and we figured we were very close to that number, surely close enough to have successful automation. It was only after going through the “Assess” phase that we learned from our subject matter experts that we had missed a key part of the automation.

We learned during the Assess phase that not all Condition Reports are the same. In fact, there was a drastically asynchronous cost associated with wrong predictions. Overall accuracy was important, but what would make or break the project was the percentage of “Conditions Adverse to Quality” that we incorrectly classified as Not Conditions Adverse to Quality. The reverse error (classifying Not Adverse to Quality as Adverse to Quality) had a cost associated with it – we might end up performing some unnecessary paperwork. But get a few high-profile errors the other way, and we would potentially miss safety-impacting conditions, undermining regulatory trust in the automation system and the CAP Program as a whole.

As a result, we made some fundamental changes to the AI models, as well as the automation system that would eventually be implemented. The AI models were trained to be more conscious of higher-profile errors, and the automation system would take into consideration “confidence” levels of predictions, with a more conservative bias. A thorough Assessment phase reduced the risk of adverse consequences and ensured any pitfalls were detected and mitigated prior to implementation.

3) Recommend

After the Assess phase, Recommendation begins. The Recommend phase typically involves providing the AI results to the manual task performers in real-time, but not automating any of their decisions. This stage is often very low risk – if the AI system is wrong or incorrect there is someone manually reviewing and correcting the errors, preventing any major inaccuracies. This is also the first stage that realizes delivered value of an AI-driven automation system.

Increased manual efficiency is often recognized as a benefit in the Recommend phase. In the majority of cases, it is physically faster for someone to perform a review of the AI’s output and make small corrections versus working the decision task from start to finish. Paired with a proper Human/AI interface, the cognitive load, manual data entry, and the number of keystrokes/mouse clicks are drastically reduced. This helps drive human efficiencies that translate to cost savings.

The Recommendation phase also permits the capture of metrics tracking how well the automation system is performing under real-world use. This is absolutely critical if partial or complete automation is desired. By running in a recommendation setup, exact data about performance can be gathered and analyzed to help improve system performance and gain a deeper objective understanding of your automation risk. This is important for deciding how to proceed with any partial automation while providing evidence to help convince those skeptical of automation.

This stage may last as short as a few weeks, or as long as several years. If the automation system needs additional training data and tweaking, the recommendation phase provides a long runway for doing so in a safe manner. Since the system is already delivering value, there is relieved pressure to reach additional levels of automation. On some projects, this phase may provide enough ROI on its own that stakeholders no longer feel the need to take on additional risk with partial or complete automation.

4) Semi-Automate

The next step in DARSA is “Semi-Automate”. This is the first stage to both fully realize the benefits and risks of automation. This phase is characterized by true automation of a task – but only for a subset of the total tasks performed.

The metrics gathered in the Recommend phase play a key role here, as they can inform which parts of the task are acceptable for automation. As the system encounters different inputs and situations, total system confidence will vary. Based on this confidence, among other metrics, automation can be implemented as a graded approach. Low-risk, high-confidence tasks are usually automated first, and after the system continues to learn, and stakeholder confidence is improved, higher-risk automations can be turned on.

For example, take a system intended to automate the planning and scheduling functions required for Nuclear Work Management. Such a system would begin to partially automate the scheduling of work activities that have low safety and operational impacts, and the planning of repetitive activities that have little historical deviation in the execution work steps. These activities are low-risk (if something goes wrong there are minimal consequences) and high-confidence (the AI has lots of previous examples with defined conditions).

During semi-automation, it is prudent to still have a manual review of a portion of automated tasks to monitor model performance, as well as provide additional training data. Without manual review, there is no longer a “ground truth” for items that have been automated. This makes it challenging to know whether the system is working well! Additionally, AI performance may begin to stagnate without the inclusion of new training examples, similar to how human performance may stagnate without new learnings and experiences.

5) Automate

The final phase that every automation system aims to achieve: complete automation. This phase is characterized by the complete automation of all tasks planned in the scope of the project. The system has been running for long enough and has gathered enough data to prove that there is no human involvement necessary. From this point forward, the only costs associated with the task are the costs associated with running and maintaining the system. Complete automation is more common with tasks that have a lower overall level of risk, yet require a lot of manual effort without automation. The most common example of this in the Nuclear Industry today is automated Corrective Action Program trend coding.

It is expected to jump back and forth between partial and complete automation at some frequency. A common case where this can occur is when the automated task or decision is changed, and the automation system hasn’t learned what those changes are. There will need to be some amount of manual intervention until the system learns the new changes, and full automation can be turned back on. An example of this would be in a “trend-coding” automation system when the “codes” or “tags” applied to data are altered.

Start Using DARSA

DARSA provides a proven roadmap to designing, building, implementing, and iterating an AI-driven automation system. Using this process, organizations embarking on the development of new automation systems can deliver maximum value with minimal risk, using a methodology appropriate for modern AI in practical automation applications.

Visit to learn more about how Nuclearn uses DARSA to help Nuclear Power Plants achieve AI-driven automation.