Surrogate-Assisted Reliability, Optimization, and Parameter Back-Analysis in Geotechnical Engineering: A State-of-the-Art Synthesis

Beneath the cities we walk on and the mountains we tunnel through lies a world of soil and rock that refuses to behave like a simple, uniform block. Engineers must design foundations, slopes, and tunnels that can withstand the unpredictable nature of the ground, where a patch of clay might be soft in one spot and hard in the next, or where water pressure shifts in ways that are difficult to predict. To ensure safety, they rely on complex computer simulations that act as digital laboratories, testing how a structure will hold up against these hidden forces. However, these simulations are incredibly slow and expensive to run. A single test of a tunnel's stability might take hours on a supercomputer, and to be truly safe, engineers need to run thousands of these tests to account for every possible variation in the ground. This creates a bottleneck: the need for certainty clashes with the limits of computing power, leaving many critical decisions stuck in a state of uncertainty or relying on overly cautious, expensive designs.

A new review by researchers Serges Mendomo Meye and Pieride Mabe Fogang charts a path through this difficulty by examining how the field is moving away from old, rigid calculation methods toward a smarter, more adaptive approach. The authors analyze a suite of modern tools known as surrogate models, which act as fast, lightweight stand-ins for the heavy, slow computer simulations. Instead of running the full, time-consuming physics engine for every single test, these new methods build a mathematical map based on a few key examples and then use that map to predict the outcome of millions of other scenarios in a fraction of a second. The review, covering the most recent advances up to 2026, finds that while older methods worked well for simple, predictable ground conditions, they fail when faced with the complex, jagged reality of real-world soil. The paper argues that the future lies in combining these fast maps with active learning, where the computer decides exactly where to look next, and physics-informed learning, where the computer is taught the fundamental laws of how soil moves, ensuring it never predicts something physically impossible.

The researchers begin by explaining that for decades, engineers relied on static methods to create these fast maps. They would pick a fixed set of test points, run the slow simulations, and draw a smooth curve through the results. This worked fine when the ground was uniform and the problems were simple. However, the authors show that this approach breaks down when the ground is highly variable or when the failure point is sharp and unpredictable, like a landslide that happens suddenly rather than gradually. In these complex situations, a smooth curve misses the danger zones entirely. The paper details a shift toward "active learning," a process where the computer does not just wait for data but actively seeks it out. Imagine a hiker trying to find the edge of a cliff in thick fog; a static map might just show a flat line, but an active learner would take a step, check the ground, and if it feels unstable, take another step right there to map the edge precisely. In the same way, these modern algorithms run a few initial tests, build a rough map, and then intelligently choose the next most important spot to test, focusing their energy only on the areas where failure is most likely. This allows them to find the exact boundary between safety and collapse with far fewer computer runs than before.

The review categorizes six main types of these fast mapping tools, explaining how each has evolved. Older methods, like polynomial response surfaces, are like drawing a straight line through a jagged mountain range; they are too simple to capture the sharp peaks and deep valleys of real soil behavior. Newer tools, such as support vector regression and radial basis functions, are better at handling these sharp turns. But the most significant leap forward described in the paper involves "physics-informed" models. These are not just guessing based on data; they are taught the rules of physics, such as how water flows through soil or how rock deforms under pressure. By embedding these laws directly into the computer's learning process, the model can make accurate predictions even when there is very little data available. This is crucial for real-world projects where sensors might be sparse or where the ground conditions are unique and have never been seen before. The authors show that these physics-guided models can predict the stability of a tunnel or the settlement of a building with high accuracy, even when the input data is incomplete.

The paper then applies these concepts to three critical areas of geotechnical engineering. First, in reliability analysis, the new methods allow engineers to calculate the probability of failure with a level of precision that was previously impossible. Instead of guessing that a slope is safe because it passed a few tests, the computer can simulate billions of scenarios in seconds, identifying the tiny, rare combinations of conditions that could lead to a collapse. Second, in design optimization, these tools help engineers find the perfect balance between cost and safety. They can rapidly test thousands of different foundation designs, adjusting the amount of concrete or steel, to find the most efficient solution that still meets strict safety codes. Finally, the review highlights the potential for real-time "digital twins." In this scenario, sensors on a construction site feed live data into the fast model, which instantly updates the understanding of the ground conditions. If the ground starts to shift in an unexpected way, the system can immediately recalculate the risks and suggest changes to the construction plan, turning the traditional method of "observe and react" into a proactive, real-time safety system.

Despite these advances, the authors are careful to point out that the field still faces significant hurdles. The complexity of the ground means that as the number of variables increases, the difficulty of creating an accurate map grows exponentially, a problem known as the "curse of dimensionality." The paper notes that while the new methods are powerful, they can still be misled by noisy data from sensors or errors in the computer simulations themselves. If the input data is flawed, the fast map will produce a flawed prediction. The researchers emphasize that future work must focus on filtering out this noise and automating the selection of the best tools for each specific job, so that engineers do not have to manually tune complex settings. They also call for models that can handle multiple types of failure at once, rather than looking at just one risk in isolation, to better reflect the interconnected nature of real-world infrastructure.

Ultimately, this synthesis serves as a roadmap for the next generation of geotechnical engineering. It confirms that the era of relying solely on slow, brute-force computer simulations is ending, replaced by a more intelligent, adaptive approach that combines speed with physical accuracy. The authors conclude that by bridging the gap between historical mechanical principles and modern data-driven techniques, the industry can move toward infrastructure that is not only safer and more resilient but also more efficient to build. The path forward involves refining these fast models to handle the messiness of the real world, automating their use so they are accessible to all engineers, and integrating them into live construction sites to create a dynamic, responsive relationship between the built environment and the ground it rests upon.