Harnessing the Power of AI & Machine Learning in Geotechnical Engineering

All over the news, you have heard about Chat GPT and AI. You also know how AI could improve productivity and increase workers' output in all industries. This is true. In fact, this is an underestimation of what AI can do. Most focus is on how Large Language Models (LLMs) can help. However, LLMs like Chat GPT rely on something far more powerful than what you hear in the news. They rely on a specific algorithm called Neural Networks.

Understanding Neural Networks

A Neural Network is a computer programming algorithm that learns data in a manner similar to how biological neurons do. In fact, all that Chat GPT does is predict the next word. Its Neural Network tells it what the next word would be based on the previous ones. This is not to take away from how useful and powerful Chat GPT is. Its Neural Network is so massive in size that it requires a vast array of NVIDIA GPUs to run and train on data.

The Neural Network algorithm can help in many other ways. It can learn any function, any data, or anything in ways the human brain cannot. It can learn and make accurate predictions without the need to formulate a rigorous mathematical formulation for the observed data. Universal Approximation Theory states that neural networks with sufficient layers and neurons can learn any continuous function at any desired accuracy. This makes Neural Networks a very powerful regression tool.

Applications of Neural Networks

In this article, the uses of Neural Networks and Large Language Models are illustrated. Useful applications are discussed in detail so that geotechnical engineers can apply them to their workflow to ensure they can get the most out of this powerful technology. Other professionals from different specializations can also learn from this article and apply the insights to their workflow. What is illustrated here has already been applied in Geo-Mechanica in consultations, software development, and day-to-day tasks.

Large Language Models

Most people have probably used Chat GPT and asked it different questions. At work, people might ask it to write, review, and summarize emails, articles, reports, or presentations. When they get home, they might ask the same technology how many calories are in their food, how to improve their lifestyle, or what the recipe is for making a delicious apple pie. Financial analysts would go as far as using it to analyze stocks, financial statements, and market news summaries.

Scientists would ask it for the best regression model to use, what can be inferred from data, and how to improve experiments. Now, most likely, the reader here is a Geotechnical Engineer. So, Geotechnical Engineers can ask it to summarize soil reports, suggest a foundation design, and propose a soil improvement program. These are just some of the things a user of LLMs would ask, and there are many others.

Limitations of General LLMs

The problem with general LLMs like Chat GPT is that they are trained on all the internet. The language is general; the information can be corrupted and inaccurate. There is no guarantee that the prompt would yield a correct output. This, in addition to hallucinations, can make LLMs' output less reliable in many cases. Hallucinations refer to the tendency of LLMs to give very weird and incorrect outputs when the context is lengthy. Give Chat GPT a one-line question, and it can provide a very good answer. However, give it a 100-page report, and you might be surprised by how many errors are in the summary it provides. You might even find information that was not in the report to begin with. In fact, any LLM user should always review the output of an LLM before making any decisions based on it.

Retrieval Augmented Generation (RAG) LLMs

This is where Retrieval Augmented Generation (RAG) LLMs are superior to general LLMs. RAG models are a type of LLM that provides output based on the data they are trained on and given to them. When asked to give information outside their training data, they do not provide an answer. This significantly reduces the chances of hallucinations. One such tool is called Notebook LM, developed by Google. In Notebook LM, you can give it many types of documents, such as websites, PDFs, or even a YouTube video. You can then ask it to summarize, review, and suggest improvements based only on the input you have given it.

In Geo-Mechanica, we created a RAG LLM called Dan, trained on many geotechnical design codes and books. It has been very useful for Geo-Mechanica in retrieving information from these codes, which is far faster and more efficient than scrolling through PDFs or flipping pages. In addition, Dan can do many things for Geotechnical Engineers, such as suggesting designs, foundation layouts, and soil profile analyses. A free version is available on our website (lower right corner). Go ahead and ask it and tell us what it told you. Not only does Dan provide geotechnical help, but it also assists with office work, such as searching internal documents, reports, and data. The more experience we gain, the more experience it gets. You can also contact us for access to the full powerful version. We can also develop custom Dans tailored to your company and needs.

Example of Using Dan

Let’s see an example of using Dan. It was given an image showing a real soil profile. It was then told that the soil would be excavated and supported. The following prompt was used:

Prompt: What is the best excavation support method based on the soil profile given? Are there any concerns I should look closely into? And what is the typical number of struts I should use in this case?

Response:

Based on the soil profile shown in your diagram, I can provide recommendations for excavation support methods and key considerations.

Recommended Excavation Support Method

For your soil profile with alternating clay (CL) and silty sand (SM) layers and an excavation depth of approximately 12.3m, I recommend a braced excavation method with the following components:

Primary Support System:

• Diaphragm wall or secant pile wall for the retaining structure

• Horizontal struts with wales for internal bracing

• Center posts to support the strut system

  
  

Alternative Options:

• Soldier pile and lagging with horizontal struts (if groundwater is not a major concern)

• Sheet pile walls with bracing (though less suitable for this depth and mixed soil conditions)

  
  

Number of Struts Required

For a 12.3m deep excavation, you should typically plan for 3-4 levels of struts:

• Strut spacing = 3-4 meters

• Typical arrangement:

- 1st level: 2-3m below ground surface

- 2nd level: 5-6m depth

- 3rd level: 8-9m depth

- 4th level: 11-12m depth (if needed)

The exact number depends on:

• Strut load capacity

• Wall stiffness

• Soil strength parameters

• Allowable wall deflections

  
  

Critical Concerns to Address

1. Mixed Soil Conditions

   The alternating CL and SM layers create several challenges:

2. Differential settlements between layers

3. Varying earth pressures at different depths

5. Potential for piping at sand layer interfaces




6. Groundwater Management

   This is likely your most critical concern:

7. Sandy layers (SM) may have high permeability

8. Dewatering will be essential to prevent:

9. Sand boiling

11. Drained shear strength of clay




12. Wall Penetration Depth

    The wall should penetrate sufficiently below the excavation level:

13. Minimum 0.7H to 1.0H below excavation base

15. Check both push-in stability and basal heave

Design Recommendations

• Earth Pressure Distribution

Use apparent earth pressure diagrams for strut design:

- For mixed conditions, consider the more conservative clay pressure distribution

- Apply appropriate factors of safety (typically 1.5-2.0)

• Construction Sequence

1. Install retaining wall first

2. Excavate in stages (2-3m lifts maximum)

3. Install struts immediately after each excavation stage

4. Implement dewatering system before excavation begins

Monitoring Requirements

• Inclinometers in the retaining wall

• Settlement monitoring of adjacent structures

• Piezometers for groundwater monitoring

• Strut load monitoring

  
  

The mixed soil profile requires careful attention to groundwater control and staged construction to ensure stability throughout the excavation process.

You can try it yourself. It is very smart. When asked how many struts and how thick a diaphragm wall should be, it replied with 3 to 4 struts and that the wall should be 0.6 m thick. This was a very good estimate compared to the case study given to it.

Conclusion

To wrap it up, what LLMs can do is to retain and reason with experience and knowledge, whether public or internal to a company, in a very human way to help shape future decision-making. Dan will become more experienced as we gain more experience. It will become more knowledgeable about the ways of Geo-Mechanica as a company. It will also become more knowledgeable in Geotechnical Engineering as more and more reports, analyses, and data are fed into its Neural Network.

There are multiple levels of using LLMs for Geotechnical Engineers, beginning from day-to-day work (Level 1) to providing insights and suggestions on designs and decisions (Level 4).

Levels of Engagement with Dan

Here are some Level 1 cases for Dan and RAG LLMs in general:

• Reply to emails

• Review reports for linguistic errors

• Summarize reports, emails, and give you suggestions for next steps based on those summaries

• Schedule your work based on importance and deadlines

  
  

Level 2 is more specialized, and here are some Level 2 tasks:

• Show it a soil report or a geologic model and ask what interesting parts to look for or what hazards are present.

• Ask it to bring you case studies from literature similar to a project you are working on.

• Ask it to provide you with relevant local code guidelines for a specific soil profile, retaining wall, or a foundation for a building.

  
  

In Level 3, Dan can get more involved:

• Ask it to review a design you came up with and check if it meets relevant codes or if it is suitable for the soil.

• Ask it to review your calculations and check for any errors or inconsistencies.

• Ask if the calculation method used is suitable for the problem and if there are better methods available.

• Ask if the adopted soil profile is good and does not ignore important features.

  
  

In Level 4, the AI can perform more complex tasks, make calculations itself, and generate a report that you can then review and approve. No decision should be left solely to AI without human insights, as these models can make errors. However, to improve the output of the model, previous calculations and reports can be provided to it so that it has a solid basis to begin with. Large tasks should be divided; for example, do not ask Dan to create a full report that is 40 to 50 pages long. Instead, break it down into subtasks. Task 1 could be soil profiling. Task 2 would involve giving it the same soil profile it made for foundation suggestions or excavation support. Task 3 could include bearing capacity and settlement calculations for foundations or wall and ground deformation for excavations.

Neural Networks in Geotechnical Engineering

Artificial Neural Networks are the algorithm behind LLMs, but their applications extend far beyond that. They can be very powerful in creating models and making predictions in many areas of science and engineering. In Geotechnical Engineering, a Neural Network can be trained to predict lateral wall deformation based on case studies or a firm's own measured data. The Neural Network can then be used to design future excavation support. This is just one example. Decision-making processes can also be quantified and embedded into a Neural Network to facilitate future decision-making. Many more applications can benefit from this powerful algorithm.

In Geo-Mechanica, we built ML Lab, a Neural Network building software that can be used to train neural networks on all sorts of data. It is a free tool and is available for site subscribers here. A step-by-step video tutorial on how to use ML Lab is provided in this video. Try it. It is very easy to use. You are literally 3-6 clicks away from building your first neural network.

Suggestions for Utilizing Neural Networks

Here are some suggestions on how Neural Networks can be utilized as an analysis and design tool for geotechnical engineers:

• Train Neural Networks on different foundation designs you have come up with from your previous work. This is very useful when using piles and piled rafts, as the number of variables involved is very large. Use the trained model on new designs to come up with design alternatives much faster.

• Similar to the example in the video shown above, you can train a Neural Network on case studies of excavations.

• Your corporate decisions that have led to desired outcomes and some that have led to undesirable outcomes can be quantified. Train a Neural Network that can help future decision-makers in your firm make informed and better decisions. After many decisions, the neural network can become so smart that decision-making can be automated and just reviewed by senior staff.

• If you work with Finite Element Codes and run analyses, you know how time-consuming it can be to model complex problems. Train a Neural Network on the results of previous Finite Element analyses so that it can inform future model building.

  
  

Step-by-Step Example on Training a Neural Network on ML Lab

ML Lab is free and can be downloaded from here. Once downloaded, install it and run the program. In this example, a Neural Network is trained on data measuring diaphragm wall maximum deformation as a result of excavation. The variables used to inform the network include 13 variables (e.g., excavation depth, number of struts, etc.). The output of the neural network is the maximum wall deformation, as shown in Fig. 1, and a sample of data points is shown in Table 1. You can also watch the example from the video here.

Table 1 showing samples of data points from excavation data

| Construction method | Total excavation stages | Excavation width (m) | Final excavation depth (m) | Wall length (m) | Wall thickness (m) | EI (MN-m²/m) | No. of struts | Avg spacing | Depth of hard stratum (m) | ΣHₑ /H wall | Ei / σ' | Su / σ' | δₕₘ/Hₑ (%) |

|---------------------|------------------------|-----------------------|----------------------------|------------------|--------------------|---------------|----------------|-------------|---------------------------|---------------|---------|---------|--------------|

| 1 | 7 | 33.4 | 18.5 | 31 | 0.8 | 918 | 6 | 3.08 | 31 | 0.87 | 650 | 0.1 | 0.33 |

| 1 | 7 | 33.4 | 18.5 | 31 | 0.8 | 918 | 6 | 3.08 | 31 | 0.87 | 650 | 0.1 | 0.33 |

Couple of things are needed before plugging in those data into ML Lab. First, you need to make sure to normalize the output values by dividing them over the maximum value to ensure the output falls between zero and 1. This is because the neural network is better trained when the predicted values are not larger than 1. There are other issues that may arise in data, which are further discussed in the manual and in video 1. Next, the data must be in CSV format. If you are using a spreadsheet, you can manipulate your data as you wish, but the final data should be exported to a CSV file.

Now open ML Lab and import your data file. Insert the number of input columns and the output column (only one output is allowed) as shown in Fig. 2.

In the Neural Network hyperparameter tab page (Fig. 3), you can adjust the number of layers, the number of neurons per layer, the learning rate, the epochs, and many more. Once you adjust those, click on the Neural Network button just below the file path entry and wait for training to finish.

If you want to use the neural network, you will find the file path at which it was saved in the log. Copy the path, check the "use the neural network for prediction" checkbox, and fill in the path to the prediction file and the neural network.

Conclusion

Machine Learning and AI are transforming how everything is done in all specialties. They can increase productivity and minimize the time needed for many tasks. LLMs and RAG LLMs use cases have been presented and categorized based on the depth of involvement of the LLM.

Most people think of Chat GPT as the only thing related to AI. However, AI and machine learning are much more than that. Chat GPT is an LLM that uses Neural Networks to perform its functions. Neural Networks are powerful algorithms that can be used in many ways. Examples of use cases have been given in this article, and an example of using Geo-Mechanica ML Lab software to train a Neural Network was provided as a video tutorial.

Disclaimer

AI/ML outputs (including LLMs/RAG/neural networks) are provided for informational purposes only and are not professional engineering or legal advice. Geotechnical decisions must be based on project-specific data and independently verified using accepted methods and applicable codes. Any designs, reports, or recommendations must be reviewed and, where required, sealed by a qualified licensed geotechnical engineer. Do not input confidential or proprietary project information into AI tools unless authorized and compliant with your policies and contracts. The authors disclaim liability for outcomes resulting from reliance on AI-generated content. AI outputs must be checked against local codes, project constraints, and professional judgment.