How civil engineers can use their data to create verifiable AI systems

We stand at a pivotal time in the civil engineering sector.

The next three to five years promise transformative shifts, thanks to advancements in technologies like large language models (LLMs), vector databases, and knowledge graphs.

These are paving the way for innovative, knowledge-driven engineering systems, much like experienced engineers leveraging their expertise to tackle new challenges.

Emerging technologies in knowledge-driven AI

Here are some of the technologies driving the knowledge revolution:

Large language models (LLMs)

These AI models, trained on vast internet-scale datasets, exhibit an emergent reasoning ability. Capable of processing text, images, audio, and video, LLMs, such as ChatGPT, can generate human-like responses in these formats.

Fine-tuned LLMs

When LLMs receive additional training on specific domain data (e.g., company technical reports), they become adept at engineering-related tasks, such as interpreting documents, aiding in design, and generating reports.

Vector embeddings

These are numerical sequences representing complex data (like images or text), enabling the handling of diverse information types.

Vector databases

By storing vector embeddings, these databases facilitate efficient searching and retrieval of semantically similar items, such as linking 'advanced construction materials' to 'graphene-enhanced concrete'.

Knowledge graphs

These map all company data, interlinking various elements and their relationships, thereby forming a comprehensive, navigable repository or archive of collective knowledge and experience.

Welcome to the knowledge economy

In this era, knowledge is often equated with time - in how long it takes to apply it.

However, the technologies mentioned earlier enable us to apply gathered knowledge to new problems.

This is easy to get a feel of – just drop a few sections of technical text from different sources into ChatGPT and ask it questions that require knowledge from all to answer.

The key is considering the LLM as the reasoning engine rather than the knowledge source.

Then, the challenge becomes providing relevant information for the LLM to reason over in the context of your query and how to verify the output.

This is the basis for a knowledge-driven engineering reasoning system that can provide correct and verifiable responses to complex, domain-specific queries.

How to mitigate hallucinations

While LLMs can generate incorrect statements – often referred to as hallucinations – employing them as reasoning engines enables the model to use the knowledge base to form responses.

When combined with fact-based reasoning, it’s possible to significantly reduce hallucinations, enabling the LLM to complete accurate, domain-specific tasks.

The company knowledge base: making ALL information available

If we take this concept to its logical conclusion, we can imagine that every project and company document forms part of this dynamic and evolving knowledge base that can help solve complex engineering problems.

Information is accessible in real-time and can be queried to find projects, documents and solutions that can be used as knowledge to solve problems at hand.

Selecting the right data architecture is therefore crucial.

There are generally two formats that are suitable for this: vector databases and knowledge graphs.

A comparison is provided in a table at the link below, including knowledge graphs with vector search, which combines both techniques in a single database.

View the comparison table

The figure below shows how these technologies lay the foundation for generative engineering systems.

It should be noted that there’s a lot of low hanging fruit – just getting a generative engineering roadmap and conducting a data assessment will be hugely beneficial.

The risk of not doing so is that when such technologies are launched, if you’re not already near the start line, you may never be able to catch up due to the exponential effect of properly applied AI.

Generative engineering example

Consider the task of drafting a detailed technical report on "Microbially-Induced Corrosion's Impact on Steel Piling in Oil Terminal Harbors".

This subject is highly specialised, extending beyond the general capabilities of a standard LLM.

The complexity lies not only in the technical nature of the topic, but also in integrating differing yet relevant pieces of information to form a coherent understanding.

However, with a knowledge base enriched with a wide variety of materials like technical reports, research papers, and detailed articles specific to this topic, an advanced AI system could successfully synthesise this diverse information.

This synthesis is comparable to how a seasoned engineer leverages their expertise to merge different pieces of data, conduct thorough analyses, and develop practical solutions.

Such an AI, when specifically tuned with this niche domain data, can replicate this sophisticated process of cognitive synthesis.

It could adeptly analyse patterns, make connections, and draw conclusions from the knowledge available.

This system's capability to deliver well-informed and precise solutions is enhanced as its knowledge base is updated with new projects, research articles and case studies.

An entirely new industry

Ultimately, a comprehensive, curated knowledge base coupled with a sophisticated and multi-layered AI system will create an entirely new industry of generative engineering.

There are, however, many technical and non-technical challenges that will need to be overcome before generative engineering systems become widespread in the industry.

Keep an eye out for my next blog, where I’ll discuss these challenges in more detail and outline some of the ways engineering firms can start to integrate the awesome power of AI into their core business.