Machine Learning Rebuilds Missing Sonic Logs and Keeps Rock Classification on Track

Deep beneath the seafloor of Australia’s Browse Basin, the rock record tells a story that energy explorers, and those weighing the region for carbon dioxide storage, desperately need to read. One of the most important instruments for that reading is the compressional sonic log, known as DTCO, which measures how fast sound waves travel through rock surrounding a borehole. Yet sonic logs are notoriously fragile: fractures, poor borehole conditions, tool eccentricity, and signal attenuation can wipe out long stretches of data. A new study has now tested, in unusually rigorous fashion, what happens when half of that critical log vanishes in the most damaging way possible, and whether machine learning can stitch it back together without corrupting the geological picture.

The research, conducted on openly accessible well log data from the National Offshore Petroleum Information Management System, is notable for how honestly it simulates data loss. Most benchmarks delete values at random, a scenario scientists call missing completely at random. Real drilling rarely behaves that politely. When sonic tools fail, they tend to fail on the extreme values, in fractured zones where waves scatter and first arrivals become hard to detect. The researchers therefore built a missing-not-at-random masking scheme: within each of six lithology classes, values were removed from the lowest and highest 20 percent of the DTCO distribution, at missingness levels escalating from 10 to 50 percent. This self-masking design is deliberately cruel to imputation algorithms, because the very values that vanish are the ones least represented in what remains.

The dataset itself demanded careful curation. Seven wells yielded 65,190 measurement rows, but the raw logs were riddled with gaps, including missingness rates above 60 percent for DTCO and nearly 82 percent for the shear sonic log. After filtering, six wells and six lithology classes survived: claystone, siltstone, silty claystone, sandstone, volcanic rock, and calcareous claystone, totaling 6,811 rows. Each sample carried seven log measurements, from gamma ray and bulk density to neutron porosity and two resistivity curves, alongside a lithology label drawn from the public archive without reinterpretation.

Four strategies for handling the artificially deleted values went head to head. Listwise deletion simply discards every row containing a missing value. Mean imputation replaces each gap with a single average number. Multivariate imputation by chained equations, or MICE, iteratively predicts each incomplete variable from the others using linear regression. The fourth contender was a Light Gradient Boosting Machine regressor, an ensemble method that builds trees sequentially, each one correcting the errors of the last, and which can capture the nonlinear web of relationships among petrophysical curves.

The results were stark. Mean imputation failed outright under the simulated missingness, producing root mean squared errors of roughly 18.4 to 18.7 microseconds per foot and negative coefficients of determination between minus 0.257 and minus 0.326, meaning it performed worse than simply guessing the mean of the missing rows. The problem is structural: masking strips away the tails of each distribution, while a single mean describes only the center. Listwise deletion fared no better as a strategy, discarding half of all data at the 50 percent masking level, including precisely the tail samples that are hardest to classify.

The two model-based methods, by contrast, held their ground. Both MICE and the LGBM regressor maintained R-squared values above 0.90 at every masking level. From 10 to 30 percent missingness, the gradient boosting approach was clearly superior, with RMSE between 3.216 and 3.351 microseconds per foot compared with 4.023 to 4.452 for MICE. At 40 and 50 percent, the gap closed and MICE edged slightly ahead at the highest level. Parity plots revealed why the boosting method’s errors grew: its predictions plateau at both extremes, around 97 to 98 microseconds per foot at the upper tail and 61 to 65 at the lower tail. Tree-based models cannot extrapolate beyond the range of their training data, and once the tails were almost entirely masked, the imputer had to estimate values it had barely seen.

The study’s most consequential move came next. Rather than stopping at reconstruction statistics, the researchers fed the imputed DTCO values, together with the six intact logs, into five classifiers: a decision tree, k-nearest neighbors, random forest, XGBoost, and an LGBM classifier. The evaluation used leave-one-well-out validation, holding out an entire well as a blind test in each of six folds, a design that prevents the information leakage that inflates scores when random splits mix correlated samples from the same borehole. Class imbalance was handled with SMOTE oversampling applied only to training data, and results were pooled across all folds and averaged over six random seeds.

Here the findings turn genuinely surprising. Even with half the sonic log deleted and machine-reconstructed, lithology classification barely budged. Under the unmasked baseline, random forest led with an accuracy of 0.606 and a macro F1 of 0.406. At 50 percent masking, its accuracy was 0.612 and macro F1 0.404, while the largest change for any classifier was a drop of roughly 0.03 in macro F1 for the decision tree. The authors are careful about interpretation: because the other six logs remained complete and alternative imputation methods were not carried into the classification stage, the stability cannot be attributed solely to imputation quality. Still, the result challenges the assumption that reconstruction error maps directly onto downstream damage, echoing broader benchmark studies showing that the lowest-error imputer does not always yield the best classifier.

The absolute performance levels deserve honest framing. Accuracy between 0.54 and 0.61, with balanced accuracy and macro F1 near 0.37 to 0.41, reflects a hard problem: claystone dominates the dataset, no well contains all six classes, calcareous claystone appears in only two wells, and sonic slowness overlaps across lithologies because porosity, clay content, compaction, and cementation all shape it. Group-based validation also yields more conservative estimates than random splitting, which is precisely why the authors argue it should be standard for models destined for new wells.

The practical takeaways are measured. Missing sonic values need not doom a dataset to deletion; MICE and gradient boosting imputation offer viable alternatives when companion logs exist, and mean imputation should be avoided for tail-dependent missingness. Method choice can follow severity, with the boosting regressor preferable below 40 percent missingness and either method defensible beyond. And imputed DTCO remained usable for classification up to the highest level tested. The authors caution that all six wells come from a single basin, that shear sonic was available as a predictor despite co-missing with DTCO at a rate of 99.92 percent in the raw data, and that 50 percent masking is a study boundary, not a universal limit. Extending the framework to multiple simultaneously missing logs, other basins, and field-derived missingness patterns remains future work. For now, the study offers subsurface energy assessment something rare: an end-to-end accounting of what data loss really costs.

Subject of Research: Machine learning imputation of missing compressional sonic log data and its impact on lithology classification in the Browse Basin

Article Title: Machine learning-based DTCO log imputation and lithology classification under a simulated MNAR missingness mechanism: A Browse Basin case study

Article References: Garini, S. A., Shiddiqi, A. M., Prabuwono, A. S., Putro, H. P., Studiawan, H., Pratomo, B. A., Ariyani, N. F., & Fahriani, N. (2026). Machine learning-based DTCO log imputation and lithology classification under a simulated MNAR missingness mechanism: A Browse Basin case study. Results in Engineering, 32, Article 113603. https://doi.org/10.1016/j.rineng.2026.113603

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113603

Keywords: well logs, machine learning, imputation, MNAR missingness, lithology classification, Browse Basin, sonic log, gradient boosting, MICE, leave-one-well-out, random forest, CO2 storage