Deep Learning in Geotechnical Engineering: A Critical Assessment
Research Paper#Geotechnical Engineering, Deep Learning, Physics-Informed Neural Networks (PINNs), Deep Operator Networks (DeepONet)🔬 Research|Analyzed: Jan 3, 2026 17:14•
Published: Dec 30, 2025 17:23
•1 min read
•ArXivAnalysis
This paper critically assesses the application of deep learning methods (PINNs, DeepONet, GNS) in geotechnical engineering, comparing their performance against traditional solvers. It highlights significant drawbacks in terms of speed, accuracy, and generalizability, particularly for extrapolation. The study emphasizes the importance of using appropriate methods based on the specific problem and data characteristics, advocating for traditional solvers and automatic differentiation where applicable.
Key Takeaways
- •Deep learning methods like PINNs and DeepONet are often significantly slower and less accurate than traditional solvers for geotechnical problems.
- •Extrapolation beyond the training data envelope is a major challenge for these methods.
- •Automatic differentiation through traditional solvers is recommended for inverse problems.
- •Site-based cross-validation is crucial to account for spatial autocorrelation.
- •Neural networks should be reserved for problems where traditional solvers are genuinely expensive and predictions remain within the training envelope.
Reference / Citation
View Original"PINNs run 90,000 times slower than finite difference with larger errors."