Here’s what I mean. Imagine an engineer embarking on a new project equipped with only a computer-aided design (CAD) file. Traditionally, gaining insights about the project location's demographics and environmental aspects would entail a cumbersome external research and data analysis process.
AI tools integrated into CAD environments simplify the design process. For instance, when an engineer selects a point on a map, middleware communicates with AI services via an application programming interface to provide pertinent information within seconds. Large datasets such as large language models—a deep learning algorithm that summarizes, translates, predicts, or generates human-sounding text—are accessible to engineers in real time, notably within their design environments, with minimal programming or added software engineering. Further, if an agency fine-tunes its large language model, appropriate stakeholders could access highly focused information.
With improved efficiency and accuracy in the preliminary design stages, engineers can make informed decisions swiftly, ensuring their designs are innovative and contextually appropriate. Such a system saves time and enriches the design process with a depth of insight previously unattainable in such a brief period.
AI also helps with another type of problem involving higher-order mathematical analysis. My passion for cycling provides an example.
Example of the Cyclist
An engineer who dreams of taking on the challenging 745-mile Paris-Brest-Paris cycling route faces a logistical problem: determining the optimal cycling hours per day to complete the ride in 90 hours while ensuring a minimum of eight hours of rest daily. The bicyclist expects to average 14 mph during the first day, with their projected speed decreasing by one mph each subsequent day.
Typically, an engineer solves the above problem by setting up a linear programming model with decision variables, constraints, data, and objective functions and then uses specialized software to determine the optimal solution. AI significantly changes the process.
The engineer selects the start and end points of the ride within CAD software on a map and, instead of setting up a linear programming model, describes the model in natural language, just as I wrote it here. The middleware, which bridges the communication between two discrete software systems, communicates with AI services, giving it the start and end coordinates and the problem statement.
AI algorithms process the input and constraints, and the engineer receives an optimized schedule within minutes. The ease with which the engineer arrives at the solution underscores the transformative power of AI in turning intricate challenges into manageable tasks.
Low Civil Engineering Risk Tolerance
Can we accept the results of AI systems unquestioningly? Of course not. Our industry, correctly, has a low risk tolerance. We still need to adhere to the expected rigor of our quality management system because it is a good ethical practice and mandated in Executive Order 13960 (Trustworthy AI) and other laws being adopted across the US.
In the first example involving the design process, AI algorithms return information about the project location, but how do engineers ensure the veracity of the information? How do we know, for example, that everything AI turns up about a part of Richmond, Va.,—its population, road system, water resources—is accurate? We can check it by triangulation, which involves using at least three diverse sources or types of data to validate findings. Engineers can validate the results by reviewing more than three sources or types of data, especially for complex and critical studies.
But when it comes to determining the optimal number of miles to bicycle each day on my Paris to Brest to Paris trip, more is needed. At a minimum, we need to generate three categories of data: one that will produce solutions, one that will not, and another that will yield errors. We use algorithms to create synthetic data, serving as a stand-in for test data to validate and train machine learning models about my trip.
The synthetic data used in testing allows us to verify and document the solutions against the expected results. For instance, if an AI system provides solutions from the two categories of synthetic data with the same results, it alerts us to a flaw in the solution.
AI is here and will be an integral part of engineering’s future. But we must apply the same level of quality management, for ethical and legal reasons, to maintain the rigor and standard of care that all engineering requires.
Anand Stephen, P.E., digital delivery leader in the roadway business group of engineer Gannett Fleming, has been instrumental in advancing digital delivery and transformation at several firms. He can be reached at astephen@gfnet.com.
