190-6 A Multi-Basin Open-Source Benchmark Comparing Physics-Informed and Conditional Tabular Generative Adversarial Network Synthetic Well Log Generators under Bayesian Optimization

Abstract:

Pore pressure prediction is critical for safe well planning, yet machine learning models designed to automate it remain constrained by the scarcity of labeled training data. Synthetic well log generation has emerged as a promising alternative, with two dominant paradigms: physics-informed generators that encode petrophysical knowledge, and data-driven models such as Conditional Tabular Generative Adversarial Networks (CTGAN) that learn directly from empirical well distributions. In this work, we introduce the Physics-Informed Synthetic Well Generator (PISWG), a modular forward-modeling framework that chains classical petrophysical relationships into a unified synthetic data generator. PISWG is calibrated using Tree-structured Parzen Estimation (TPE) with a softmax-weighted multi-objective fidelity loss. We benchmark PISWG against an identically tuned CTGAN baseline across five geologically distinct regions. To the best of our knowledge, no prior work has systematically compared these approaches on a multi-basin benchmark under a controlled hyperparameter budget while incorporating both distributional and downstream evaluations.

Our results reveal a consistent but incomplete picture. CTGAN achieves substantially better distributional fidelity across all tested regions, attaining lower Jensen–Shannon divergence (0.197 vs. 0.372 mean JSD) and higher downstream Train-Synthetic-Test-Real (TSTR) performance across six regression baselines. In contrast, PISWG provides approximately three times higher Distance-to-Closest-Record (DCR) protection, calibrates nearly an order of magnitude faster (31 vs. 339 minutes average wall time), and exposes interpretable physical parameters that can be inspected by domain experts.

However, neither approach achieves reliable generalization under a strict leave-one-well-out protocol on the Eastern Potwar Basin. Even with up to 20% real-data augmentation, both synthetic-data approaches and real-only baselines yield predominantly negative R² values on several held-out wells. These findings expose a fidelity–utility gap: strong distributional similarity does not necessarily translate into robust downstream predictive performance in low-data scientific machine learning settings.

Geological Society of America Abstracts with Programs. Vol. 58, No. 5, 2026

doi: 10.1130/abs/2026AM-15611

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.