Machine Learning Types and Their Infrastructure Use Cases

AI and Machine learning is a complex field with numerous models and varied techniques. Understanding these different types and the problems that each is suited to solving is key to unlocking their potential in Infrastructure.

The built environment is essential for supporting the stability and safety of modern life. The planning, design, construction, maintenance, and operation of our civil infrastructure is a massive undertaking, with demands increasing as populations grow and the strain on time, people, and resources approaching breaking point. Managing each stage of the infrastructure lifecycle requires deep expertise and careful allocation of resources, informed by ever-increasing quantities and forms of data. This is where AI has a significant role to play in alleviating the burden on engineers, maximising process and project efficiencies, reducing costs, and saving time. 

First, however, those responsible for enacting AI adoption in infrastructure must understand the types of AI available and map them to the use cases they are best suited to maximise their impact. There is a wide range of AI and machine learning models, and understanding the differences between each and figuring out what they are best suited for can be a challenge. Below, we have outlined some of the most widely used types of AI and identified some of the use cases in Infrastructure. 

For more insight into the potential for AI in Infrastructure, download our whitepaper: AI for Civil Engineers: What You Need to Know to Build the Future.

What are some of the different types of AI and Machine Learning models?

AI is a field of computer science in which machines perform tasks that had previously been possible exclusively by human intelligence. However, the term AI is not well-defined and can cover anything from simple rule-based systems to complex deep learning models. 

Rule-based systems 

Rule-based systems run on predefined rules to make decisions and usually follow “if-then” statements. They are a basic type of model that uses logical reasoning and decision-making based on predefined rules and conditions. They are optimal for repetitive, rule-based tasks where information, or data, is collected and evaluated against expert rules to achieve some desired outcome. For example, a rule-based system could enable a council to optimise road closure times for infrastructure repair to minimise disruption and the associated costs. A subject expert would provide specific rules, and the model would output Task Order recommendations optimised to meet the demands and rules. 

Machine learning 

Machine learning refers to systems that can learn and improve from relevant datasets without being explicitly programmed. It is a subset of AI that involves training algorithms to recognise patterns and make decisions based on data. Instead of being explicitly programmed, machine learning models learn from examples and improve their performance over time. Machine learning consists of three main learning paradigms: supervised learning, unsupervised learning, and reinforcement learning. 

1. Supervised learning: 

Supervised learning is a type of machine learning where a model is trained using labelled data, meaning each input comes with a corresponding output. The model learns to map inputs to outputs by finding patterns in the training data, allowing it to make predictions for new, unseen data. Common supervised learning algorithms include: 

Classification: Classification algorithms predict discrete outputs, often referred to as class labels. Models created using classification algorithms exploit the relationship between the data describing an entity or instance and the class label associated with that entity. The relationship can then be used to predict the correct label of a given input data set with an unknown label. An example of a challenge well-suited to a classification model could be looking through construction documentation and flagging excessive or unjustified Change Orders that deviate from industry norms as A) Suspicious or B) Legitimate, thus saving valuable time and resources for the individuals who would otherwise be tasked with these investigations. 

Regression: Regression algorithms predict continuous outputs. Models created using regression algorithms learn a function that describes the relationship between one or more independent variables in the input data and a response, dependent, or target variable. Regression models could add real value to condition intelligence by taking an image of a defect and, on a logarithmic scale from 1 to 5, giving the defect a damage severity rating that could be used to prioritise inspection and maintenance planning.

The model is trained using labelled data, so each input comes with a corresponding output.

Commonly used types of supervised machine learning models include: 

Linear models: Linear models are the simplest type of machine learning model. They search for the best-fitting line through a set of data points, making them useful for understanding and predicting how changes in one variable affect another when the relationship between the two is consistent. 

Decision trees: Decision trees are supervised learning algorithms used for both classification and regression tasks. They have a hierarchical structure consisting of a root node, branches, internal nodes, and leaf nodes. Decision trees are similar to expert systems in that they can be interpreted as if-then rules that result in a final outcome. Unlike rule-based systems, though, decision trees learn the rules from labelled data rather than depending on an expert to define the rules in advance. 

Neural networks: A neural network is a computational model inspired by the structure and functioning of the human brain, consisting of layers of interconnected nodes (neurons). It is designed to recognise patterns, learn from data, and make predictions or classifications, making it a key building block in many machine learning and deep learning applications. Neural networks are highly effective for many tasks that could add real value to infrastructure where complex relationships in the data need to be captured. Examples include image recognition to help improve defect detection and natural language processing to help ingest and interrogate text data from inspection reports to enable better and faster decision-making. 

Deep neural networks: Deep learning refers to a particular type of modelling within machine learning whereby multi-layered neural networks are used to solve tasks. These deep neural networks can automatically learn representations from raw data. However, they often need millions of parameters, and as such, they are not easily interpretable. Nevertheless, deep learning models have driven significant advancements in AI due to their ability to handle high-dimensional data and achieve state-of-the-art results in various applications. 

Time series forecasting: Time series forecasting is the process of using historical data to predict future values, where time is a key feature of the data. The model will output a time-based prediction for a given time stamp. For example, based on the history of the average daily price of construction materials like steel, coupled with additional market indicators over the past year, a time series forecasting model could be used to estimate the steel price for a specific day in the future. 

Convolutional neural networks: Convolutional neural networks are an example of deep neural networks. Convolutional neural networks are a type of artificial neural network for deep learning that is often used for image and video-based learning, where spatial relationships between pixels in images can be used to identify patterns, known as feature learning, making them highly effective at tasks such as object detection and classification. A convolutional neural network could be built to analyse video feeds and camera images from construction sites, identifying unsafe operator behaviours and hazardous environments to improve safety and reduce accidents. 

Alternatively, a convolutional neural network could be built to identify different defect types - such as cracks, corrosion, spalling, and reinforcement bars - identified on various assets and prioritise maintenance of these assets based on the defect types and their anticipated deterioration pathways. This kind of insight could drastically improve condition intelligence and optimise the management of an entire portfolio of assets. 

2. Unsupervised learning:

Unsupervised learning models are trained on unlabelled data, so the model itself finds patterns in the data rather than a human providing the labels to guide it. An example of unsupervised learning is clustering models. 

Clustering models: These are machine learning techniques that aim to find similarities within data to identify different groups, or “clusters”, that exhibit different characteristics without requiring classification labels. The characteristics that separate data belonging to different clusters can be further examined to determine how they differ and can be used as a proxy for classification labels. A potentially valuable infrastructure-specific use case for this kind of model could be to cluster a set of unlabelled defect images into groups of images with similar defects. Using an approach like this could separate images with cracks from images with spalling. 

The model is trained on unlabelled data, so it finds patterns in the data without guidance from a human.

3. Reinforcement learning 

Reinforcement learning models use techniques that train software to make decisions to achieve the most optimal results. They mimic the trial-and-error learning process that humans use to achieve their goals. The model uses a reward system to determine how optimal each action is with respect to the desired objective. The reward is sometimes a human providing feedback on the model output (saying whether the output is correct or not). As such, the model can iteratively improve its behaviour. 

From a civil infrastructure perspective, a reinforcement learning model could be deployed to forecast optimal bridge maintenance schedules. The model’s reward system might look to minimise the risk of bridge failure due to insufficient maintenance while lowering the cost of maintenance. The users of that model could then provide feedback back to the model based on the bridge’s true condition following inspections suggested by the model. 

The model uses a reward system to determine how optimal each action is with respect to the desired objective.

Generative AI and Large Language Models 

Generative AI and Large Language Models are a subset of deep learning models, and they have exploded into the public consciousness with models like OpenAI’s GPT-4, Google’s Gemini, Mistral AI’s Mixtral, and, most recently, DeepSeek-R1. Generative AI refers to a methodology which learns relationships across disparate data domains and the concepts underneath them from incredibly large, broad sets of data. Generative AI can subsequently create a wide variety of outputs, such as images, text, audio, and datasets. Large Language Models are very similar; however, they specifically generate natural language. More generally, generative models learn the patterns and structure of data and then create new data with similar characteristics. 

While impressive, the true value proposition and range of enterprise use cases of these models are still very unclear, so they should not be considered catch-all solutions. These models usually remain too generic to generate real business value, and their factual inaccuracies, or “hallucinations”, do not inspire confidence in a high-stakes environment like civil infrastructure.

AI Is Not a One-Size-Fits-All Solution

Although AI’s potential is undeniable, and there is a real opportunity to use it to elevate our civil infrastructure management, merely adopting AI will not reap automatic rewards. AI is immensely complex and varied, as are the problems in infrastructure that it could help solve. The model types listed above are not the only ones available, but if civil engineers can gain a basic understanding of the types of AI that are most widely applicable in their work, it will help them to identify the problems in their work that the technology could help solve. 

This is one of the crucial steps in AI adoption in infrastructure and is key to the sector's success as AI and machine learning become more widespread in the years to come.

To find out what civil engineers need to know about AI and Machine learning to build the future, download our whitepaper: AI for Civil Engineers: What You Need to Know to Build the Future.