A young engineer may be amazed by the seasoned reviewer who picks out a deficient element in their design, without carrying out calculations or running a computer model. The experienced engineer is tapping into the wealth of results of their own prior analyses and the many designs they have come across. Now imagine this knowledge acquisition being enhanced thousands of times over. This is what an AI can be trained to do.
This Insight article explores how a machine learning (ML) bot can be expected to first assist, then perhaps replace the human expert and go beyond.
What is AI to be Trained on?
For millennia, builders learned the trade through their own trials and errors as well as those of their ancestors. There were no theories and no formulas that explained how it works. Galileo is believed to have been the first to attempt to develop a beam theory, but it took until 1750 AD for a valid theory to be formed by Leonhard Euler and Daniel Bernoulli. Ever since, engineers have evolved an enormous array of techniques to rationally idealize, model and simulate the behavior of structures. Parallel to this theoretical approach were equally important empirical formulations that rely on testing programs of various materials and systems. Some of those are now enshrined in building codes. So, will AI learn the theories and the formulas, or will it go back to the old way of learning patterns that work, but with mammoth digital power? Let’s reflect on that.
A Generative Pre-trained Transformer (GPT) can be, and is already in some practices, trained to spit out information from codes and standards in response to plain language queries. Some GPTs can be focused on one field while others such as ChatGPT, Claude 3 and Google Gemini are generic and constantly scour the internet for more data. They are, though, not so precise or reliable.
Supervised ML algorithms use brain-like processes that are called neural networks. They establish patterns and make interpretations, augmented by probabilistic reasoning, to respond to queries within a learnt dataset. Unsupervised learning is another approach where no human guidance is used to assist with the learning. Rather, the algorithm learns patterns exclusively from unlabeled data. The key to either approach is the extent and applicability of the dataset. So, what datasets are available for structural analysis and design ML algorithms?
The First Digital Transformation in Structural Engineering
The advent of computer power has transformed the structural engineering profession so profoundly over the past four decades. Key to this revolution is the finite element analysis (FEA). This is an approach that discretizes complex systems into small enough elements that can be attributed linear or nonlinear constitutive relationships. The resulting models can then be statically or dynamically analyzed. Today, every engineering office uses these tools that fifty years ago were confined to universities and research institutions. Will these be given up, run in parallel with AI or will there be different transformational shifts?
AI Can Learn a Thing or Two from FEA Simulations
As a way around the limitations of the size of the available datasets and the infinite possibilities of design, AI will do what engineers always do – approximate. Surrogate models are trained using FEA simulations to establish patterns that allow the algorithms to offer fast (even if approximate) results without the need to run expensive simulations.
While it’s early days, surrogate models have already been applied to optimize reinforced concrete floors, to mimic nonlinear dynamic analysis and to conduct performance-based design. In one case a surrogate model was trained on a dataset generated using results from nonlinear time-history analyses of thirty reinforced concrete buildings with different structural dynamic characteristics, which are subjected to sixty-five real ground motions. The resulting AI circumvents the need to perform such elaborate analysis where an assessed building falls within these parameters. Some of these attempts may be simplistic but telltales are there of much more to come.
What about Structural Design?
On the empirical front, a ML algorithm, XGBoost, has been used to access the raw experimental data from various testing programs that were used to generate an empirical formula of the code and provide design parameters that are even more consistent with the data than the code formula. In fact, theoretically, such a ML algorithm can acquire data from experimental publications in real time and constantly improve on design parameters — a feat that codes cannot keep up with. Or, it must be said, do not want to keep up with so promptly, for various regulatory, legal and professional reasons. This gives a hint of some of the nuances and limitations involved in design.
Design software packages are plentiful. These are based on traditional automations rather than AI, as are a whole host of tools that facilitate the modeling, analysis, design and production of construction drawings. So which frontiers will AI need to conquer?
Structural design practice is mostly based on linear elastic models which are known to have limitations. The alternatives that model materials and systems better are expensive and time consuming. As mentioned, surrogate models can be trained to offer fast and efficient answers that account for the finer material and system idealizations. Another field being explored by AI researchers is design optimization. In today’s practice, professional engineers compare a few options that they perceive, from their own experience, to be feasible and economic. They optimize, to some extent, their weight or dimensions. AI algorithms have been developed, albeit at a research level, to explore geometries, connectivities, materials and other parameters to arrive at optimal solutions. To date, these are mostly confined to individual structural subassemblies.
More robust AI models are needed, and will inevitably be available at some point, that can optimize design while considering structural typologies, connectivity, constructability, sustainability and even aesthetics. For the latter, research on AI in the field of the intangible is well underway, where in one case ML was trained through responses of architects to pick the most aesthetic of originally generated designs. Not only would this contribute to design automation but it promises to overcome human biases in the process. In time, it is quite conceivable that AI will ultimately produce design alternatives and forms beyond human cognitive abilities. The next question then is: who will sign on such designs and take on the associated liability?
Previous Insight
