SlumpGuard: AI & Data-Driven Instability Control

• SlumpGuard is a family of decision support systems combining AI and physics-based models to predict and control slump in concrete and geotechnical contexts.
• Key methodologies include video-based analysis with YOLOv8/ResNet-3D, geometric semantic genetic programming, and thixotropy-aware modeling, achieving high precision and speed.
• Implications extend to enhanced quality assurance in concrete placement, early-warning for slope failures, and active control through simulation and sensor fusion.

Searching arXiv for "SlumpGuard" and closely related papers to ground the article in published work. SlumpGuard is a research label applied to several data-driven and physics-informed decision-support systems concerned with predicting, monitoring, or controlling “slump” or instability phenomena in civil and geotechnical engineering. In its most direct and published sense, it denotes an AI-powered, real-time system for automated concrete slump prediction via video analysis (Kim et al., 14 Jul 2025). In related technical syntheses, the same name is used for a practical engine that forecasts the slump of recycled aggregate concrete (RAC) from mix-design variables using geometric semantic genetic programming (GSGP) (Xu et al., 2017), and for early-warning or control frameworks for slope failure, thixotropy-aware workability degradation, and active correction of granular-foundation tilt (Hongran et al., 2017, Tordesillas et al., 2021, Wilkes et al., 2021, Martínez-Ortíz et al., 2024, Roy et al., 2024). Across these uses, SlumpGuard refers not to a single universal algorithm but to a family of operational pipelines that couple sensing, predictive modeling, and decision logic to support quality assurance or hazard management.

1. Definition and scope

The most specific use of the term appears in "SlumpGuard: An AI-Powered Real-Time System for Automated Concrete Slump Prediction via Video Analysis" (Kim et al., 14 Jul 2025). There, SlumpGuard is defined as an automated, real-time, video-based system that infers concrete slump directly from the visual dynamics of discharge from truck chutes. The task is framed as follows: given a video sequence $x_{1:T}$ of concrete discharging from one of two truck chutes, the system predicts a categorical slump class $\hat{y}$ corresponding to actual slump $y$ in millimeters (Kim et al., 14 Jul 2025).

A second, conceptually distinct use appears in the synthesis built around "Geometric Semantic Genetic Programming Algorithm and Slump Prediction" (Xu et al., 2017). In that formulation, SlumpGuard is conceived as a practical, data-driven engine to forecast and control the slump of recycled aggregate concrete before batching, using a trained GSGP model over mix-design variables such as cement, fly ash, water, sand, stone, recycled aggregate, water reducer, and total mass (Xu et al., 2017).

The name is also extended in several technical adaptations to geotechnical monitoring and active control problems. These include a locked-segment slope-instability warning framework derived from a physical model of rock-slope failure (Hongran et al., 2017), a spatiotemporal slope-stability analytics framework derived from network-flow analysis of ground-motion data (Tordesillas et al., 2021), a thixotropy-aware slump-flow decision-support concept for Tremie concrete (Wilkes et al., 2021), hybrid FEM-to-MPM landslide or slump-runout simulation guidance (Sordo et al., 2022), an active granular-foundation tilt-control concept based on an inflatable bladder (Martínez-Ortíz et al., 2024), and slope-stability screening based on physically consistent slip-surface parametrization or unsaturated upper-bound limit analysis (Lalicata et al., 2024, Roy et al., 2024).

This breadth suggests that “SlumpGuard” functions as an umbrella designation for monitoring and intervention architectures centered on predictive inference under material variability, incomplete observability, and operational constraints.

2. Concrete slump monitoring by video analysis

In the 2025 video-based system, SlumpGuard addresses the limitations of manual slump testing, which is described as labor-intensive, subject to operator variability, and usually sampling only a small portion of a batch (Kim et al., 14 Jul 2025). The system therefore aims at full-batch, real-time quality assurance through automated chute-video interpretation.

The pipeline comprises five stages. First, a ZED 2i stereo camera captures both chutes simultaneously at combined resolution $3840 \times 1080$, consisting of two synchronized $1920 \times 1080$ streams. Second, a YOLOv8 oriented object detector identifies each chute and produces both rotated and upright bounding boxes. Once detections remain consistent for more than 8 consecutive frames, the region of interest is locked by averaging box parameters and detector inference is paused to reduce compute. Third, sparse optical flow by Lucas–Kanade tracks the motion at the chute’s box center, and the discharge start time $t_0$ is detected when the tracked center crosses the bottom edge of the rotated bounding box. Fourth, from $t_0$, $N$ consecutive frames from the corresponding upright-cropped chute patch are processed by a ResNet-3D classifier, with the deployed backbone being $R(2+1)D$-18. Fifth, the predicted slump class is compared against the ordered slump range and out-of-spec deliveries are flagged (Kim et al., 14 Jul 2025).

The system discretizes continuous slump into five categorical intervals: under 150 mm, 150–180 mm, 180–210 mm, 210–240 mm, and over 240 mm. The mid-range class widths are aligned to the $\pm 30$ mm tolerance of KS F 4009, whereas the extreme ranges do not use explicit margins because of rarity (Kim et al., 14 Jul 2025).

The empirical performance reported for this pipeline is strong. Chute detection achieved $\hat{y}$0, Precision $\hat{y}$1, and an average speed of 9.0 ms per frame, with detector inference paused after RoI locking (Kim et al., 14 Jul 2025). Placement timing and active-chute identification are reported as consistently above 90%, with aggregate statements reporting “average accuracy over 95%,” though the right chute was affected by hopper-cover shadow and very high slump could weaken optical-flow signal (Kim et al., 14 Jul 2025). For slump-class prediction, $\hat{y}$2-18 achieved test accuracy $\hat{y}$3 and macro $\hat{y}$4, outperforming TimeSFormer, MC3-18, and R3D-18 on the reported split (Kim et al., 14 Jul 2025).

The training configuration is explicitly documented. Clips contain 16 frames sampled at interval 2 with temporal extension factor 2; augmentation includes random horizontal flip, ColorJitter, MixUp with $\hat{y}$5, CutMix with $\hat{y}$6, label smoothing $\hat{y}$7, and resize ratio sampled from $\hat{y}$8. Optimization uses AdamW with learning rate $\hat{y}$9, weight decay $y$0, total batch size 128, 10 epochs, and a OneCycle scheduler with cosine and peak learning rate $y$1 (Kim et al., 14 Jul 2025).

3. Mix-design prediction for recycled aggregate concrete

A separate SlumpGuard formulation is based on the GSGP regression model proposed in "Geometric Semantic Genetic Programming Algorithm and Slump Prediction" (Xu et al., 2017). The motivating problem is that recycled concrete shows complex composition and high variability in fresh properties because recycled aggregates carry adhered mortar, have higher porosity, diverse particle size distributions, and more fines than natural aggregates. Within that setting, conventional methods for forecasting slump scarcely obtain satisfactory results, and the paper motivates an intelligent prediction model tailored to nonlinear behavior (Xu et al., 2017).

The RAC dataset contains 34 mixes. The first 28 samples were used for training and the remaining 6 for testing (Xu et al., 2017). The terminator set comprises eight practical mix-design variables: cement, fly ash, water, sand, stone, recycled aggregate, water reducer, and total mass. Although one earlier sentence mentions “six input units,” the explicit terminator set comprises eight variables, and the SlumpGuard synthesis follows the eight-variable specification (Xu et al., 2017).

The GSGP formulation uses arithmetic operators $y$2 as function set and the eight variables as terminators. The original fitness is defined as the sum of absolute errors over instances,

$y$3

where $y$4 is the prediction of individual $y$5 for instance $y$6, and $y$7 is the experimental slump. RMSE is also used for error reporting,

$y$8

and the paper assesses linear association using a Pearson coefficient $y$9 (Xu et al., 2017).

The main hyperparameters reported are an initial population size of 500, maximum generations of 50, and mutation step $3840 \times 1080$0 (Xu et al., 2017). The geometric semantic operators are

$3840 \times 1080$1

for crossover and

$3840 \times 1080$2

for mutation, where the random functions take values in $3840 \times 1080$3 (Xu et al., 2017).

Reported performance indicates that the model exhibits strong agreement with experimental slump values. The paper reports relative errors generally below 5% on the test set and a Pearson correlation coefficient of $3840 \times 1080$4, approximately 97.65% (Xu et al., 2017). Over 50 independent runs, the RMSE interquartile range was 0.0892 for GSGP, compared with 1.9933 for LSSVM and 3.8651 for STGP, with Wilcoxon rank-sum p-values of $3840 \times 1080$5, $3840 \times 1080$6, and $3840 \times 1080$7, respectively (Xu et al., 2017). One reported test-set pair gives computed slump 130.9 mm versus experimental slump 125.2 mm, an absolute error of 5.7 mm, about 4.6% (Xu et al., 2017).

In deployment guidance, the synthesis describes a workflow in which a candidate mix design is input, slump is predicted, the result is evaluated against a target, and small changes in water, superplasticizer, fines, or recycled aggregate content are explored before batching (Xu et al., 2017). Some of these steps, such as explicit use of tournament selection with elitism, cross-validation, optional complexity penalties, additional metrics such as MAE and $3840 \times 1080$8, and ensemble uncertainty bands, are presented as deployment practices rather than as claims of the original paper (Xu et al., 2017). This suggests a distinction between the published GSGP model and later operational interpretations of “SlumpGuard” built around it.

4. Thixotropy-aware workability control

Another SlumpGuard concept is derived from the study of Tremie concrete thixotropy in "Investigating the Thixotropic Behaviour of Tremie Concrete Using the Slump-flow Test and the Material Point Method" (Wilkes et al., 2021). Here the emphasis is not batch-to-batch visual inspection or symbolic regression, but the history dependence of fresh-concrete workability.

The paper develops a thixotropic model integrating the Papanastasiou-Bingham model with thixotropy equations in the Material Point Method framework. The constitutive structure uses a structural parameter $3840 \times 1080$9 that increases during rest and decays during flow. In the form stated in the synthesis, the dynamic-yield-stress relation is

$1920 \times 1080$0

with rest-state build-up approximated by

$1920 \times 1080$1

and shear-induced breakdown modeled as

$1920 \times 1080$2

for a constant shear-rate segment (Wilkes et al., 2021).

The integrated Papanastasiou–Roussel constitutive law is written as

$1920 \times 1080$3

with reported choices $1920 \times 1080$4 and $1920 \times 1080$5 in the simulations (Wilkes et al., 2021). The virtual slump-flow simulation uses approximately 17,000 material points, a background grid of approximately 256,000 nodes and approximately 219,000 hexahedral GIMP elements, a cone lift speed of 0.15 m/s, friction coefficient calibrated in the range 0.35–0.45, and explicit time step $1920 \times 1080$6 s (Wilkes et al., 2021).

Validation is reported for three Tremie mixes. Mix A has $1920 \times 1080$7 Pa, $1920 \times 1080$8 Pa·s, $1920 \times 1080$9 Pa/s, and physical spread values $t_0$0 mm and $t_0$1 mm. Mix B has $t_0$2 Pa, $t_0$3 Pa·s, $t_0$4 Pa/s, and physical spreads 550 mm and 530 mm. Mix C has $t_0$5 Pa, $t_0$6 Pa·s, $t_0$7 Pa/s, and physical spreads 500 mm and 485 mm (Wilkes et al., 2021).

The central result is that simulated $t_0$8 values after 240 s rest are approximately 94–105 mm, about 18–20% of $t_0$9, while physical $t_0$0 values under standard protocols are only 15–20 mm, about 2.6–4% of $t_0$1 (Wilkes et al., 2021). The paper interprets this discrepancy as evidence that the physical slump-flow test can mask true thixotropic loss because of factors such as bleed water under an unsealed cone and operator-dependent lift speed (Wilkes et al., 2021). In the SlumpGuard framing, this becomes a decision-support problem: track elapsed rest time, update $t_0$2, estimate current yield stress, and predict expected slump-flow response before workability loss compromises placement (Wilkes et al., 2021).

This suggests a broader sense in which SlumpGuard denotes not just a predictor of slump magnitude but a state-estimation layer for concrete rheology under temporal degradation.

5. Geotechnical reinterpretations of SlumpGuard

Outside concrete workability, the name is repurposed for slope instability analytics and related geotechnical decision support. In the synthesis of "A physical model predicting instability of rock slopes with locked segments along a potential slip surface" (Hongran et al., 2017), SlumpGuard denotes an early-warning framework that operationalizes a physical relation between the displacement at peak strength and the displacement at the onset of volume dilation. The key practical rule is

$t_0$3

and for a slope with $t_0$4 locked segments,

$t_0$5

The ratio 1.48 is described as approximately constant for Weibull shape parameter $t_0$6 (Hongran et al., 2017). In this reinterpretation, SlumpGuard detects the first sustained acceleration, estimates the number of locked segments, and predicts the final critical displacement threshold (Hongran et al., 2017).

A second geotechnical use draws on "Spatiotemporal slope stability analytics for failure estimation (SSSAFE): linking radar data to the fundamental dynamics of granular failure" (Tordesillas et al., 2021). There, SlumpGuard is described as a monitoring system that uses network flow theory and mesoscience to infer emergent kinematic clusters from surface displacement data. The core capacity function is

$t_0$7

and the framework tracks minimum cuts, silhouette score, normalized mutual information, and inverse-velocity forecasts to identify the geometry, location, and time of failure (Tordesillas et al., 2021). This is a conceptually different object from concrete slump monitoring, but it shares the same operational pattern: streaming observations, latent-state inference, anomaly detection, and thresholded warning logic.

Further geotechnical extensions include efficient limit-equilibrium search using physically consistent slip-line parametrization (Lalicata et al., 2024) and upper-bound unsaturated slope-stability screening under surcharge and pseudo-static seismic loading (Roy et al., 2024). In the former, a hybrid coarse-grid plus Nelder–Mead search around Bishop Simplified analysis achieves 80–92% time reduction relative to a dense center–radius grid and average accuracy improvement of about 5% (Lalicata et al., 2024). In the latter, suction-stress-based effective stress is coupled to steady vertical flow, log-spiral failure kinematics, surcharge, and seismic coefficients, with chart-based guidance for stability number under evaporation, no-flow, infiltration, and saturated conditions (Roy et al., 2024).

These geotechnical applications are not equivalent to the concrete-focused SlumpGuard system of 2025. A plausible implication is that the term has evolved into a reusable research name for guardrail-like monitoring architectures around slump, collapse, or instability transitions.

6. Active control and simulation-oriented extensions

The label also appears in control-oriented and simulation-oriented adaptations that extend beyond passive prediction. One such extension is based on "A smart granular intruder" (Martínez-Ortíz et al., 2024). There, a cylindrical intruder equipped with an inflatable latex bladder detects angular velocity using an MPU-6050 and inflates the bladder when $t_0$8 exceeds a threshold. The chosen optimal threshold is 70 deg/s, within an observed robust plateau from 60 to 80 deg/s, and the bladder reaches final diameter of approximately 1.4 cm in approximately 0.1 s (Martínez-Ortíz et al., 2024). With two ballast units, passive rotation is approximately $t_0$9 and lateral drift approximately 1.0 cm, while the smart intruder suppresses both within approximately 0.2–0.3 s after release (Martínez-Ortíz et al., 2024). In the SlumpGuard-oriented synthesis, this mechanism is generalized to “smart foundations” that counteract tilt by rerouting force chains through selective inflation (Martínez-Ortíz et al., 2024).

Another extension is based on "Hybrid Finite Element and Material Point Method to Simulate Granular Column Collapse from Failure Initiation to Runout" (Sordo et al., 2022). That work is not itself a SlumpGuard paper, but its synthesis frames it as a simulation core for slump or landslide progression. The hybrid FEM→MPM workflow transfers coordinates, velocities, and stresses from FEM to MPM particles after initiation but before severe mesh distortion, with tested transfer times of 0, 0.5, 1.0, 2.0, and 3.0 s (Sordo et al., 2022). For a granular column with aspect ratio $N$0, early transfers yield final normalized runout around 1.6, close to the empirical value 1.53, whereas later transfer degrades accuracy (Sordo et al., 2022). This use places SlumpGuard in the role of a simulation-backed decision engine for failure initiation and runout rather than a field-deployed monitoring product.

These examples show that SlumpGuard has been associated not only with measurement and forecasting but also with intervention logic and numerical experimentation. This suggests that the unifying theme is a control loop around instability-relevant observables, rather than a fixed domain or model class.

7. Limitations, ambiguities, and research directions

A recurring limitation across SlumpGuard formulations is domain specificity. The video-based system predicts ordered slump classes, not continuous slump values, and its reported deployment conditions do not include night operations or adverse weather (Kim et al., 14 Jul 2025). The RAC GSGP model uses only 34 mixes, does not explicitly include recycled-aggregate absorption, density, or particle-size distribution, and the slump standard is not specified in the paper (Xu et al., 2017). The thixotropy-aware slump-flow formulation depends on calibrated rheological and frictional parameters and uses a simplified single-parameter structural state $N$1 (Wilkes et al., 2021).

A second limitation is terminological ambiguity. SlumpGuard can denote at least three substantially different technical objects: a video-based chute-analysis system (Kim et al., 14 Jul 2025), a symbolic-regression engine for recycled-concrete slump (Xu et al., 2017), and a broader family of early-warning or control frameworks in slope mechanics and granular media (Hongran et al., 2017, Tordesillas et al., 2021, Lalicata et al., 2024, Roy et al., 2024, Martínez-Ortíz et al., 2024). The concrete claims associated with one version should not be transferred uncritically to another.

The literature also indicates several future directions. For the video-based system, these include continuous slump regression, multi-camera fusion, physics-informed modeling of free-surface flow and rheology, BIM integration, and uncertainty quantification (Kim et al., 14 Jul 2025). For the GSGP RAC formulation, future work includes larger datasets, broader ranges of recycled-aggregate properties, and symbolic simplification or parsimony control (Xu et al., 2017). For the thixotropy-aware framework, practical extensions include surrogate models built from offline PR-Bingham–MPM simulations and integration with scheduling, re-mixing, and admixture adjustment logic (Wilkes et al., 2021). In geotechnical applications, future work points toward uncertainty propagation, sensor fusion, non-circular or three-dimensional slip mechanisms, and tighter coupling of rapid screening methods with physics-based simulators (Tordesillas et al., 2021, Lalicata et al., 2024, Roy et al., 2024).

Taken together, these developments indicate that SlumpGuard is best understood as a cross-cutting research construct at the intersection of sensing, predictive modeling, and operational decision support. In concrete technology it targets workability assessment before or during placement; in geotechnics it is used as a label for warning, screening, or control systems for instability. The shared technical motif is the same: infer an otherwise costly or delayed state variable from structured observations, then act on that inference before the system enters an unacceptable regime.