AI Generates Physically Consistent Synthetic Rock Strength Data for Engineers

Machine-learning models promise to accelerate rock mechanics and geotechnical design, but they have long been starved of the one ingredient they need most: reliable laboratory measurements of uniaxial compressive strength, or UCS. Testing rock cores to failure is destructive, expensive, and slow, and the coordinated measurements that do exist are unevenly spread across rock types. A new study now reports a way to manufacture more of this scarce data without breaking the physics that makes it meaningful.

Researchers Kursat Kilic and Yewuhalashet Fissha, publishing in the journal AI in Civil Engineering, developed a diversity-tuned physics-informed generative adversarial network, which they call PI-GAN, designed to produce synthetic UCS, porosity, and bulk-density values for rock materials. Unlike conventional tabular data generators, which can reproduce statistical patterns while silently violating the mechanical relationships that govern how rocks actually behave, the new framework embeds rock-mechanics constraints directly into the training objective.

The model was calibrated on 944 complete records drawn from the P3 petrophysical database, an open-access repository containing coordinated UCS, porosity, and dry-density measurements. The samples were grouped into four geological categories: dense crystalline rocks (363 samples), sedimentary rocks (413), pyroclastic rocks (144), and volcanic rocks (24), a distribution that highlights the chronic imbalance problem in laboratory rock data. The trained PI-GAN then generated 944 synthetic samples, doubling the usable dataset to 1,888 records.

What sets the approach apart is its architecture. Separate generator and discriminator networks were trained for each rock category, and each generator maps a random latent vector to a synthetic triplet of UCS, porosity, and density. The generator’s loss function combines adversarial learning with an intact-rock strength-envelope penalty inspired by the Hoek–Brown failure criterion, differentiable correlation constraints for the UCS–porosity, UCS–density, and porosity–density relationships, moment and covariance matching, feature-envelope penalties, and a pairwise-distance diversity term. The authors stress that the constraint is inspired by the uniaxial Hoek–Brown condition rather than a full rock-mass implementation, because the database does not supply complete rock-mass parameters such as the Geological Strength Index or disturbance factors.

The physics constraints matter because the relationships among these variables are well established. In the real dataset, UCS decreases with porosity (a correlation of −0.598), reflecting the loss of load-bearing capacity as void space increases, and increases with density (0.566), while porosity and density are strongly inversely related (−0.895). Conventional augmentation methods such as SMOTE or CTGAN can reproduce marginal distributions yet fail to preserve these trends, which the authors argue can make statistically plausible synthetic data positively misleading for geotechnical engineering.

The validation results suggest the physics-informed approach worked. The overall UCS–porosity correlation shifted from −0.598 in the real data to −0.675 in the synthetic data, and the UCS–density correlation moved from 0.566 to 0.650, preserving the direction and rough magnitude of both trends. Across the four rock categories and three correlation relationships, 12 Fisher z-tests were performed, and 11 showed non-significant differences between real and synthetic correlations at the 0.05 level. The single significant difference involved the dense-crystalline porosity–density relationship, where the synthetic data slightly strengthened an already strong inverse trend. The average absolute correlation difference across category-specific pairs was 0.036.

Distributional and diversity checks added further nuance. The mean Kolmogorov–Smirnov statistic across 12 category–property combinations was 0.181, with values ranging from 0.110 to 0.292; several tests remained statistically significant, partly because of the larger sample size and residual differences at distribution tails, so the authors advise interpreting the KS results alongside correlation, maximum mean discrepancy, and memorization diagnostics. The mean MMD value was 0.022, indicating reasonable joint-distribution similarity. The mean pairwise-distance diversity ratio was 0.703, meaning the synthetic spread remains somewhat narrower than the full real distribution, a deliberate trade-off from the diversity tuning that improved coverage relative to an earlier, overly conservative version of the model. The near-duplicate rate was 0.0 percent across all categories, confirming the model did not simply copy training samples, and the memorization ratio averaged 0.896, with values near unity indicating synthetic samples sit at comparable distances to the real data rather than reproducing them exactly.

Strength-envelope sensitivity checks returned zero violations across tolerance levels from 10 to 50 percent of category mean UCS, supporting the choice of a 30 percent default tolerance. The authors also note an important limitation: the model is calibrated to laboratory-scale intact-rock measurements, and direct application to jointed or weathered rock masses is not recommended without additional variables representing discontinuity characteristics, in-situ stress, RQD, GSI, and disturbance factors. Extension to non-intact rock masses would require conditioning on rock-mass classification parameters and validation against field-scale data.

The intended use is deliberately modest. The synthetic records are meant to augment training data for UCS prediction models, benchmark machine-learning algorithms under controlled data-scarcity conditions, and support preliminary sensitivity analyses—not to replace laboratory testing, field characterization, or engineering judgment. Final design decisions, the authors emphasize, should remain contingent on site-specific measurements and conventional rock-engineering assessment. The training code, validation scripts, and synthetic dataset are slated for release through a public repository and the author’s GitHub account upon publication.

For a field where a single laboratory UCS test can consume a precious rock core and machine-learning models increasingly shape preliminary design decisions, the study offers a template for generating augmentation data that respect the underlying mechanics rather than merely mimicking the statistics. Whether such physics-consistent synthetic datasets become a standard supplement in geotechnical machine learning will depend on how well the approach generalizes beyond the P3 database—but the results reported here indicate that generative models can be taught to respect the laws rocks actually obey.

Subject of Research: Physics-informed generative adversarial networks for synthesizing rock strength, porosity, and density data in geotechnical applications

Article Title: Physics-informed generative adversarial networks for synthetic uniaxial compressive strength data generation in geotechnical applications

Article References: Kilic, K., & Fissha, Y. (2026). Physics-informed generative adversarial networks for synthetic uniaxial compressive strength data generation in geotechnical applications. AI in Civil Engineering, 5(1), Article 25. https://doi.org/10.1007/s43503-026-00109-6

Image Credits: AI Generated

DOI: 10.1007/s43503-026-00109-6

Keywords: PI-GAN, generative adversarial networks, uniaxial compressive strength, rock mechanics, geotechnical engineering, synthetic data, porosity, bulk density, Hoek-Brown criterion, machine learning, data augmentation, petrophysics