Deep learning models can accurately forecast the progressive tilting behavior of large gravity-type port caissons during staged construction using only a small set of initial observations, significantly outperforming traditional statistical methods. In a comprehensive numerical investigation published in the International Journal of Geo-Engineering, researchers Saeyon Kim, Jihoon Kim, and Heejung Youn formulated the prediction of caisson rotation as an early-stage limited observation forecasting problem. By utilizing stage-wise time-series data extracted from one thousand automated finite element simulations, the research team demonstrated that specialized neural network architectures successfully learned complex deformation trends across twenty-eight consecutive construction phases. Among the evaluated architectures, Temporal Convolutional Networks established the most robust predictive performance, maintaining high precision even when provided with as few as three initial construction stage measurements, while conventional autoregressive techniques failed to capture the nonlinear dynamics of structural rotation.
Gravity-type caissons represent a vital engineering solution for deep-water quay wall construction, prized across global maritime infrastructure projects for their high structural stability, favorable deformation behavior, and manageable long-term maintenance costs. These massive hollow concrete structures function primarily to transfer heavy wave actions and external operations loads directly into foundation soils, thereby securing bearing capacity and overarching stability for port terminals. However, constructing such massive structures over soft coastal deposits poses severe geotechnical stability challenges, as natural ground heterogeneity combined with asymmetric load increments frequently induces differential settlement and angular distortion. In coastal settings characterized by thick marine clay deposits, such as the Nakdong River delta region in South Korea where soft cohesive strata reach depths up to seventy meters, unconsolidated soils exhibit high natural water contents, elevated liquid limits, low shear strength, and high compressibility, making caissons highly susceptible to dangerous tilting during installation.
Structural tilting during construction impairs operational serviceability, diminishes bearing safety margins, and can precipitate catastrophic construction-stage failures or costly maintenance liabilities over the design life of maritime facilities. Consequently, engineers have historically relied on a spectrum of predictive and evaluative approaches, ranging from empirical formulas and static numerical analyses to continuous field instrumentation. Empirical methods apply historical port construction records and simplified analytical equations, such as Newmark-based displacement procedures or factor-of-safety rotational formulations, but fundamentally fail to capture nonlinear soil kinematics, detailed structural geometry, and intricate soil-structure interactions. Static finite element or finite difference simulations quantitatively evaluate displacement behavior but remain acutely sensitive to spatial variations in soil stiffness, struggling to anticipate real-time stage-wise operational deviations. Conversely, field instrumentation systems utilizing settlement gauges, inclinometers, and Global Navigation Satellite Systems provide high-precision measurements of physical displacements but remain inherently retrospective, capturing structural movement only after deformation has already materialized.
To overcome the operational limitations of retrospective monitoring and static simulations, recent advances in civil and geotechnical engineering have increasingly focused on data-driven artificial intelligence to forecast structural deformation from temporal sensor streams. Recurrent architectures such as Long Short-Term Memory networks have been deployed to anticipate consolidation settlement, tunnel deformations, and excavation wall displacements by retaining long-term sequential dependencies through specialized gating mechanisms. Concurrently, Temporal Convolutional Networks have emerged as powerful alternatives, utilizing dilated causal one-dimensional convolutions to expand receptive fields and process long input sequences in parallel without recurrent structures. Additionally, Transformer models employing multi-head self-attention mechanisms have demonstrated utility in capturing global inter-temporal correlations across multi-sensor monitoring networks for bridge deflection and deep foundation excavations. Despite these technological strides, the comparative capacity of disparate machine learning architectures to forecast subsequent caisson tilt under conditions where only initial construction stages are observed had remained largely unexplored.
To establish a rigorous comparative benchmark, the researchers developed a high-fidelity two-dimensional plane-strain finite element model within PLAXIS2D, modeled after a representative gravity-type caisson design for a South Korean port facility. The modeled reinforced concrete caisson measured 22.5 meters in height, 25.0 meters in width, and 45.4 meters in length, structured internally as a cellular assembly featuring five transverse compartments and nine longitudinal compartments. In the numerical mesh, plate elements with a unit weight of 24.0 kilonewtons per cubic meter and an elastic modulus of 1.0 multiplied by ten to the sixteenth kilopascals were assigned to approximate rigid structural behavior. The ground stratigraphy was defined with a 24.4-meter-thick soft clay layer resting above a ten-meter-thick gravel stratum, with groundwater established at an elevation head of 28.10 meters. The soft clay was modeled using the Soft Soil constitutive formulation to capture nonlinear compression, while the underlying gravel was simulated using a drained Mohr-Coulomb model within a refined fifteen-node triangular element mesh.
The numerical framework precisely reproduced the physical sequence of harbor execution by dividing construction into seven primary operational phases spanning twenty-eight discrete numerical calculation stages. Foundation stabilization commenced with Deep Cement Mixing ground improvement, followed by placement of a foundation riprap bedding layer and marine placement of the caisson body. Subsequently, internal infilling was executed across five transverse cellular chambers in incremental lifts of 1.8 to 2.0 meters up to a final internal height of 21.2 meters, generating eleven distinct infilling stages. Once filled, external backfilling was conducted in ten sequential stages along the outer right perimeter of the structure, followed by the installation of five drainage chamber covers to seal internal compartments. Construction concluded with the placement of capping concrete and front-facing armor stone placement to ensure hydraulic protection against wave actions, creating a comprehensive stage-by-stage deformation history for structural tilt extraction.
Physical tilt values were calculated at every stage from the horizontal displacement differential between the upper and lower nodes along the leftmost caisson boundary, divided by total structural height to yield rotational angular indices. Across the staged progression, distinct mechanics governed rotation: during stages one through five, asymmetric wall thickness tapering from 600 millimeters on the left to 400 millimeters on the interior produced an eccentric self-weight bias that caused initial leftward tilting. During stages six through eleven, selective infilling concentrated within right-hand cells reversed the eccentricity, initiating rightward rotation. Backfill riprap placement in stages twelve through sixteen occurred distant from the structure, generating minimal tilt, whereas stages seventeen through twenty-one placed riprap directly adjacent to the rear wall, driving pronounced rightward tilting. Minimal movements accompanied drainage cover placement in stages twenty-two through twenty-six, while final capping concrete and armor stone installations introduced secondary rotational adjustments.
Recognizing that field-treated ground exhibits substantial spatial variability from heterogeneous water contents, operational mixing disparities, and curing variations, the authors incorporated extensive material uncertainty through Latin Hypercube Sampling. Material parameters including the elastic modulus and undrained shear strength of Deep Cement Mixing ground, as well as the stiffness and interface reduction factors of heaving soil and granular riprap materials, were defined with uniform reduction ranges up to fifty percent below design values. To represent spatial heterogeneity without artificial correlation, the improved subsoil, heaving ground, and foundation riprap zones were subdivided into thirty-three, twelve, and six independent spatial subzones, respectively. Latin Hypercube Sampling stratified each parameter range into one thousand non-overlapping intervals, producing one thousand unique, systematically sampled finite element simulation cases that captured realistic variations in foundation stiffness and resulting differential settlement magnitudes.
The generated database of one thousand stage-wise tilt histories was structured into an early-stage limited observation forecasting challenge using a temporal masking protocol. A total of 150 cases were isolated as an independent test set, with the remaining 850 cases divided into 680 training and 170 validation models, which yielded 17,000 training samples when expanded across varying initial observation windows. To prevent future temporal data leakage while preserving process awareness, construction stage indices were supplied across all twenty-eight stages, while tilt values beyond designated observation thresholds of three, five, ten, fifteen, or twenty stages were masked to zero. Performance across the forecast horizon was rigorously quantified using Mean Absolute Error, Root Mean Squared Error, and the coefficient of determination, comparing a per-case fitted Auto Regressive Integrated Moving Average statistical model against Long Short-Term Memory, Temporal Convolutional Network, and encoder-only Transformer architectures.
Quantitative evaluations revealed that deep learning models consistently outperformed the statistical baseline across every tested observation condition. The Auto Regressive Integrated Moving Average model exhibited severe predictive degradation, recording negative coefficients of determination under all conditions, including an R-squared of negative 2.4645 and an absolute error of 0.0802 degrees when given three observation stages, because linear extrapolation could not anticipate sudden mechanical inflection points caused by subsequent construction activities. In stark contrast, the Temporal Convolutional Network achieved exceptional accuracy under identical three-stage inputs, achieving a Mean Absolute Error of 0.0108 degrees, a Root Mean Squared Error of 0.0124 degrees, and an R-squared of 0.9162. While Long Short-Term Memory and Transformer models yielded respectable R-squared values of 0.8728 and 0.8825 at three observation stages, their predictive errors were noticeably higher during early phases, solidifying the architectural superiority of dilated causal convolutions under severe data scarcity.
Systematic hyperparameter exploration demonstrated that structural network complexity must be carefully balanced against generalization capacity under finite training conditions. For the Long Short-Term Memory network, elevating hidden state dimensions from 32 to 128 produced steady error reductions, but further scaling to 256 yielded negligible accuracy gains alongside sporadic performance instability. Similarly, the Temporal Convolutional Network achieved optimal stability with 128 channels, although configurations with at least 32 channels already captured primary temporal dynamics effectively. For the encoder-only Transformer, a model dimension of 32 paired with a 64-unit feed-forward layer delivered the highest forecasting precision, whereas scaling model dimensions to 64 or 128 degraded average performance. When observation horizons expanded to twenty stages, minimal tilt variations during final construction phases produced very small variance denominators that resulted in negative R-squared values across all neural architectures, despite the Temporal Convolutional Network attaining tiny absolute errors of 0.0038 degrees.
An in-depth evaluation of representative cases positioned across the training distribution revealed crucial distinctions in model behavioral mechanics. When analyzing central, maximum-tilt, and minimum-tilt test samples, the Temporal Convolutional Network demonstrated remarkable stability across both high-density central distributions and unrepresented distributional tails. Conversely, the Long Short-Term Memory network exhibited pronounced errors under limited observations but rapidly gained accuracy once ten or more stages were observed. The Transformer architecture displayed a systematic inductive bias toward the training set mean, underpredicting maximum-tilt trajectories and overpredicting minimum-tilt cases under small observation windows, an effect corroborated by a conspicuously narrowed, tall prediction distribution peak across the full 150-case test set at early stages. The authors emphasized that while these simulation-derived findings confirm the analytical power of causal convolutions, practical implementation on operational harbor sites will require field sensor calibration to account for three-dimensional geometry, continuous measurement noise, and time-dependent consolidation settlement between loading stages.
Subject of Research: Deep learning-based time-series forecasting of gravity caisson tilt during staged port construction
Article Title: Deep learning-based forecasting and model comparison of caisson tilt using numerical simulation–derived time-series data
Article References: Kim, S., Kim, J., & Youn, H. (2026). Deep learning-based forecasting and model comparison of caisson tilt using numerical simulation–derived time-series data. International Journal of Geo-Engineering, 17(1), Article 35. https://doi.org/10.1186/s40703-026-00295-3
Image Credits: AI Generated
DOI: 10.1186/s40703-026-00295-3
Keywords: caisson tilt forecasting, gravity caissons, deep learning, temporal convolutional networks, long short-term memory, transformers, ARIMA, geotechnical engineering, numerical simulation, PLAXIS2D, staged construction, deep cement mixing, Latin Hypercube Sampling, time-series analysis, soft marine clay, port engineering, soil-structure interaction
Cite Scienmag News
Blake Davidson. (November 18, 2026). Deep Learning Models Forecast Gravity Caisson Tilting From Early Construction Data. Scienmag. https://scienmag.com/deep-learning-models-forecast-gravity-caisson-tilting-from-early-construction-data/
Blake Davidson. "Deep Learning Models Forecast Gravity Caisson Tilting From Early Construction Data." Scienmag, 18 November 2026, https://scienmag.com/deep-learning-models-forecast-gravity-caisson-tilting-from-early-construction-data/. Accessed 18 November 2026.
Blake Davidson. "Deep Learning Models Forecast Gravity Caisson Tilting From Early Construction Data." Scienmag. November 18, 2026. https://scienmag.com/deep-learning-models-forecast-gravity-caisson-tilting-from-early-construction-data/








