Geotechnical Engineering AI Analysis
Deep Learning Surrogate for Undrained Cyclic Response of Sands: Stability and Generalization
This research presents a Long Short-Term Memory (LSTM) framework trained exclusively on raw experimental databases from cyclic Direct Simple Shear (DSS) and Triaxial (TXC) tests. It introduces systematic data thinning and physics-guided loss weighting to address computational and physical constraints. The model robustly predicts stress-strain-pore pressure trajectories, demonstrating stability and generalization across diverse material states.
Executive Impact: Revolutionizing Soil Behavior Prediction
Leverage AI to overcome traditional model limitations in geotechnical earthquake engineering, ensuring robust, data-driven predictions for critical infrastructure projects.
Average MAE
Experimental Databases Utilized
Key Methodological Advancements
LSTM Hidden Units
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Foundations & Challenges
Understanding the cyclic response of granular soils remains a central challenge in geotechnical earthquake engineering, particularly regarding liquefaction triggering, pore-pressure generation, and long-term cyclic degradation. Foundational empirical frameworks (Ishihara 1993; Seed and Idriss 1971) and advanced constitutive models like PM4Sand (Bou-langer and Ziotopoulou 2015) have established the basis for simulating these phenomena.
However, their application is commonly accompanied by extensive, apparatus-specific calibration to capture cumulative response characteristics across diverse stress paths (Dafalias and Manzari 2004; Liao and Yang 2021; Petalas et al. 2019; Wu and Bauer 1994; Wu et al. 1996; Yang et al. 2020). These limitations have motivated the shift toward data-driven methods capable of learning soil behaviour directly from experimental records. Early artificial intelligence (AI) applications—primarily static neural networks used for pile capacity or slope stability (Banimahd et al. 2005; Ghaboussi et al. 1991; Habibagahi and Bamdad 2003; Kohestani and Hassanlourad 2016; Shahin et al. 2001; Shin and Pande 2000)—successfully modelled nonlinear corre-lations but lacked the temporal architecture to reproduce the path-dependent nature of cyclic loading.
Key Takeaway: Traditional models require extensive calibration; AI offers a data-driven alternative but often lacks temporal understanding for path-dependent cyclic loading.
LSTM Architecture & Data Engineering
The present study adopts a consistent-architecture Long Short-Term Memory (LSTM)–based sequence-learning framework to predict the undrained cyclic response of sands under varying stress paths. The framework is applied independently to each apparatus type, yielding two separately trained models—one for cyclic direct simple shear (DSS) and one for cyclic triaxial (TXC) conditions. Both models share identical architecture (single LSTM layer, U=96), optimizer (Adam, η=10−3), training horizon (120 epochs), and validation strategy (Leave-One-Test-Out cross-validation).
The novelty of the present framework lies in the strategic biasing of the optimization gradient and the numerical compression of the history window to capture the extreme nonlinearity and path-dependency of cyclic soil response. To maintain numerical stability over sequences that can exceed 104 increments in raw laboratory data, a systematic data thinning (decimation) protocol was employed. In cyclic liquefaction modelling, to prevent the global gradient from being dominated by late-stage large deformations, a physics-guided loss weighting strategy was introduced, ensuring the precise capture of early-stage pore-pressure generation.
Key Takeaway: A novel LSTM framework is proposed, featuring systematic data decimation and physics-guided loss weighting to ensure stability and accuracy, especially in early-stage pore-pressure generation.
Predictive Performance & Generalization
The predictive capability of the proposed LSTM framework is evaluated against the Ottawa sand DSS database to assess its competence in reproducing two distinct kinematic failure mechanisms: rapid flow liquefaction (high CSR) and high-cycle cyclic mobility (low CSR). The model accurately reproduces the subtle contraction of the vertical effective stress during the initial three cycles. The precise detection of liquefaction triggering (ru ≈ 0.98 at N = 3) validates the Physics-Guided Loss Weighting strategy. By explicitly amplifying the loss gradients in the pre-liquefaction domain (ru < 0.98), the optimizer effectively learns the micro-mechanical contractive tendency that precedes failure.
The generalization capability of the framework was evaluated using the undrained cyclic triaxial (TXC) database on Karlsruhe sand. This boundary condition introduces complex stress path rotation and higher asymmetric strain accumulation than simple shear. The model captures rapid pore pressure generation and subsequent collapse of mean effective stress, terminating stably at the origin without numerical oscillations. It also reproduces high-intensity cyclic mobility, including "butterfly" loops and progressive accumulation of axial strain (ratcheting) in dense sands.
Key Takeaway: The LSTM model accurately predicts complex soil behaviors like liquefaction triggering and cyclic mobility across both DSS and TXC tests, reproducing subtle early-stage pore pressure generation and stress-strain trajectories.
Understanding Prediction Boundaries
While the proposed LSTM framework demonstrates high fidelity in predicting undrained cyclic soil response across diverse loading modes, several boundary conditions and technical limitations govern its current application. A primary limitation is the divergence in accuracy between stress and strain variables. While stress-based outputs (shear, deviatoric, and mean effective stresses) remain robust through the post-liquefaction phase, strain predictions exhibit increased uncertainty following effective stress collapse.
This disparity is attributed to the strongly path-dependent nature of strain ratcheting, which is governed by micro-structural mechanisms not explicitly captured by the current scalar input representation. Specifically, fabric anisotropy was identified as the primary governor of prediction limits in dense sands, providing a clear mechanistic basis for the observed residuals in directional strain accumulation. This highlights the necessity for incorporating directional state variables or fabric tensors in future iterations.
Key Takeaway: While robust for stress predictions, strain prediction accuracy is limited by the scalar input's inability to capture fabric anisotropy, highlighting a physical boundary for current data-driven models.
~0.175
Average Mean Absolute Error (MAE) for DSS Tests (U=96, W=20, η=5e-4)
Enterprise Process Flow
Raw Experimental Data
→
Systematic Data Decimation
→
Physics-Guided Loss Weighting
→
LSTM Model Training
→
Accurate Pore-Pressure Generation
| Feature | Proposed LSTM | PM4Sand Model |
|---|---|---|
| Training Data |
|
|
| Calibration |
|
|
| Stiffness Degradation |
|
|
| Energy Dissipation |
|
|
| Physical Constraints |
|
|
| Generalization |
|
|
Case Study: Predicting Dense Sand Dilatancy & Ratcheting (TTx-22)
The LSTM model successfully reproduced high-intensity cyclic mobility, including characteristic "butterfly" loops and progressive accumulation of axial strain (ratcheting) for dense Karlsruhe sand (Dr = 82%, p₀' = 300 kPa). Despite prolonged cycling (N >> 100) and high-amplitude stress reversals, the model effectively captured density-dependent hardening, avoiding premature failure predictions common in models that fail to resolve high dilatancy.
Challenge: Accurate tracking of cumulative plastic strain and phase transformation behavior in dense sands under high cyclic demand.
Solution: The physics-guided LSTM, trained on experimental data, internalized the frictional limit and transition from contractive to dilative tendency, ensuring stability without explicit thermodynamic constraints.
Calculate Your Potential AI-Driven ROI
Estimate the significant efficiency gains and cost savings your enterprise could achieve by implementing advanced AI solutions for geotechnical analysis.
Annual Cost Savings
$0
Annual Hours Reclaimed
0
Your AI Implementation Roadmap
A phased approach to integrate deep learning surrogates into your geotechnical workflows, from data preparation to continuous monitoring.
Phase 1: Initial Data Ingestion & Preprocessing
Establish robust pipelines for ingesting and systematically thinning raw experimental data from various sources (DSS, TXC tests). Implement quality control and standardization to ensure data integrity for model training.
Phase 2: Model Training & Calibration
Train the Long Short-Term Memory (LSTM) framework using physics-guided loss weighting to prioritize critical early-stage pore-pressure generation. Optimize hyperparameters to achieve stable and generalizable predictive performance.
Phase 3: Validation & Performance Benchmarking
Rigorously validate the trained model using Leave-One-Test-Out cross-validation on unseen soil specimens. Benchmark its performance against established constitutive models like PM4Sand, identifying strengths and limitations.
Phase 4: Deployment & Continuous Monitoring
Integrate the stable constitutive surrogate into your geotechnical design practice. Implement continuous monitoring of model predictions against new experimental data to ensure long-term accuracy and identify retraining needs.
Unlock Advanced Geotechnical Prediction for Your Enterprise
Ready to integrate cutting-edge AI for more stable, accurate, and generalized soil response predictions? Schedule a personalized consultation to explore how our solutions can meet your specific needs.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat