⚡ electrical engineering

Deep Learning Characterization of Spatially Variable Ground Using Screened and Label-Conditioned Random Field Data

This study proposes a systematic framework for generating physically consistent training datasets through non-stationarity screening and Maximum Likelihood Estimation-based label verification, demonstrating that such data quality improvements significantly enhance the accuracy and generalization of Convolutional Neural Networks in estimating the scale of fluctuation for spatially variable ground.

Original authors: Ashu Singhal, Gyan Vikash

Published 2026-08-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ashu Singhal, Gyan Vikash

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the ground beneath our feet not as a solid, uniform block of concrete, but as a giant, chaotic sponge. If you poke it in one spot, it might be hard as a rock; move your finger just a few inches away, and it could be squishy mud. This is the reality of soil: it is messy, unpredictable, and changes constantly over tiny distances. Engineers call this "spatial variability." To build safe skyscrapers, bridges, and tunnels, they need to understand how far this "messiness" stretches. Is the soil changing every foot, or does it stay the same for ten feet before shifting? This distance is called the "scale of fluctuation." Think of it like the size of the waves in the ocean; some waves are tiny ripples, while others are massive swells. If an engineer guesses the wave size wrong, their building might sink or crack.

For a long time, figuring out these wave sizes was like trying to guess the weather by looking at a single cloud. It was hard, slow, and often inaccurate. Recently, scientists started using Artificial Intelligence (AI), specifically a type called a Convolutional Neural Network (CNN), to act like a super-smart detective. These AI models can look at a long strip of soil data and instantly guess the "wave size." But here's the catch: to teach an AI, you need to show it millions of examples with the correct answers. If you teach a student using a textbook full of typos, they might memorize the typos instead of the real lessons. This is exactly the problem researchers faced when trying to teach AI about soil.

In this study, Ashu Singhal and Gyan Vikash from Shiv Nadar University decided to fix the "textbook" before letting the AI read it. They discovered that the standard way of creating fake soil data for training was full of hidden traps. Sometimes, the fake soil didn't behave like real soil at all, and other times, the "answer key" (the label telling the AI what the wave size was) didn't actually match the picture the AI was looking at. It was like showing a picture of a cat but labeling it "dog."

The researchers built a new, rigorous system to clean up this data. First, they acted like a strict editor, throwing out any fake soil samples that were too weird or "non-stationary" (meaning they didn't follow the basic rules of soil physics). Next, they used a mathematical tool called Maximum Likelihood Estimation (MLE) to double-check every single sample. They asked, "Does this picture actually look like the label says it does?" If the answer was no, they fixed the label to match the picture. Finally, they grouped similar pictures together and gave them a single, clear label, making it easier for the AI to learn the patterns without getting confused by tiny, noisy differences.

They tested four different versions of their training data: the messy original version, the cleaned-up version, and two versions where they grouped the data even more carefully. When they trained their AI on the messy, original data, the computer seemed to get perfect scores on the test, but it was actually just memorizing the typos. When they switched to their new, carefully cleaned and grouped data (specifically the version with seven distinct groups), the AI learned the real rules. The true test came when they fed the AI real-world data from actual soil tests in Adelaide, Australia. The AI trained on the messy data failed miserably, guessing the soil's "wave size" with huge errors. However, the AI trained on the new, high-quality data got it right, with errors dropping by about 73%.

The main finding is that the secret to a smart AI isn't just the brain of the computer; it's the quality of the food you feed it. You can have the most powerful engine in the world, but if you put sand in the gas tank, the car won't run. By screening out bad data and ensuring the labels physically matched the soil patterns, the researchers showed that AI can reliably predict how soil behaves in the real world. This suggests that for engineers to trust AI with building safety, they must first ensure the training data is physically consistent, not just statistically large. The study didn't just find a better algorithm; it found a better way to teach the algorithm, proving that in the world of soil science, a clean dataset is just as important as a smart computer.

Technical Summary: Deep Learning Characterization of Spatially Variable Ground Using Screened and Label-Conditioned Random Field Data

Problem Statement
The spatial variability of geotechnical properties, particularly the scale of fluctuation (SoF), is critical for stochastic geotechnical analysis and reliability assessment. While Convolutional Neural Networks (CNNs) have emerged as a promising tool for estimating SoF from site investigation data, their application relies heavily on large datasets of synthetically generated soil profiles. Recent studies utilizing LU decomposition-based random field theory to generate these datasets have identified significant limitations:

Inconsistent Realization-Label Mappings: Random field realizations generated with a prescribed SoF often exhibit effective SoF values that differ substantially from the input parameters due to the stochastic nature of the generation process.
Non-Stationarity: Some generated realizations violate the second-order stationarity assumption inherent in the stochastic models, introducing physical inconsistencies.
These issues result in training datasets containing "label noise," where the assigned labels do not accurately reflect the physical characteristics of the data, potentially leading to unreliable supervised learning models that fail to generalize to field data.

Methodology
The study proposes a systematic framework to construct physically consistent training datasets for CNNs. The methodology involves the following stages:

Data Generation: A preliminary dataset of 24,000 one-dimensional random field realizations of cone tip resistance ( $q_c$ ) was generated using LU decomposition-based random field theory with Monte Carlo simulation. The inputs included a mean of 25 MPa, a coefficient of variation of 20%, and prescribed SoF values of 0.5 m, 1 m, 2 m, and 5 m.
Screening and Verification:
- Non-Stationarity Screening: Realizations violating the second-order stationarity assumption were identified and removed, reducing the dataset to approximately 20,000 stationary realizations.
- MLE-Based Relabeling: For the retained realizations, the effective SoF was estimated using Maximum Likelihood Estimation (MLE). The original prescribed labels were replaced with these MLE-derived values to ensure the labels matched the actual spatial correlation of the realization.
Statistical Conditioning:
- Outlier Removal: Realizations with MLE-estimated SoF values falling outside the statistical whisker limits (specifically $\theta > 2.03$ m) were excluded to reduce extreme variance, resulting in a "Screened Dataset" of 18,692 realizations.
- Label-Conditioning Strategy: To address the complexity of regressing continuous MLE values, the dataset was partitioned into classes based on ascending MLE-estimated SoF. Each class was assigned a representative SoF value (the arithmetic mean of the class). Two variations were tested:
  - Label-Conditioned Dataset I: Divided into 4 equal classes.
  - Label-Conditioned Dataset II: Divided into 7 equal classes (finer discretization).
CNN Architecture and Training: A 1D Convolutional Neural Network was developed with two convolutional blocks (64 filters, kernel size 5, ReLU, MaxPooling, Dropout) followed by fully connected layers (128 and 32 neurons) and a linear output layer. Hyperparameters were optimized using a Random Search strategy. The model was trained on four distinct datasets: the unprocessed Preliminary Dataset, the Screened Dataset, and the two Label-Conditioned Datasets.

Key Results
The study evaluated model performance using the coefficient of determination ( $R^2$ ), Mean Squared Error (MSE), and relative prediction error against MLE-estimated SoF values.

Synthetic Data Performance:
- The Preliminary Dataset (unprocessed) yielded the highest apparent $R^2$ (0.901 test) but exhibited a high mean relative error (0.669), indicating the model learned the inconsistent labels rather than the physical reality.
- The Screened Dataset (MLE-relabelled) showed a drop in $R^2$ (0.688 test) and an increase in error. The authors attribute this to the increased difficulty of regressing a continuous, physically consistent distribution compared to discrete, inconsistent labels.
- Label-Conditioned Dataset II (7 classes) achieved the best balance, with a test $R^2$ of 0.891 and the lowest test MSE (0.047). The mean relative error dropped to 0.290, demonstrating that reducing label variability while preserving physical consistency improves learning stability.
Field Data Validation:
- The models were validated against 11 independent CPT profiles from Keswick Clay.
- The model trained on the Preliminary Dataset systematically underestimated the SoF, with a mean relative error of 0.731 and RMSE of 0.393 m.
- The model trained on Label-Conditioned Dataset II significantly improved generalization, reducing the mean relative error to 0.197 and RMSE to 0.130 m (a ~73% and ~67% reduction, respectively). Predictions closely followed the one-to-one reference line with the MLE-estimated reference values.

Significance and Claims
The paper claims that the reliability of CNN-based characterization of spatially variable ground is not solely dependent on network architecture but is fundamentally constrained by the quality and physical consistency of the training dataset.

The study demonstrates that:

Direct use of LU decomposition-generated random fields with prescribed labels introduces significant inconsistencies that degrade model transferability to field data.
A framework integrating non-stationarity screening, MLE-based label verification, and label-conditioning strategies effectively bridges the gap between synthetic training data and physical reality.
By establishing a physically consistent realization-SoF mapping, the proposed methodology enables CNNs to generalize successfully to independent field measurements, offering a practical pathway for improving data-driven characterization of heterogeneous ground conditions.

Technical Summary: Deep Learning Characterization of Spatially Variable Ground Using Screened and Label-Conditioned Random Field Data

More like this