AI-Driven Data-Centric Computing for Large-Scale Scientific Research: Automated Data Quality Assessment, Feature Discovery, Knowledge Extraction, and Reproducible Machine Learning Pipelines

Abstract

Machine learning is now routine in scientific research, but the reliability of its results depends at least as much on the data as on the model. Label errors, missing values, gross measurement errors and duplicated records are common in scientific datasets, and they are rarely measured before a model is trained. This study presents and evaluates an open, end-to-end data-centric pipeline that combines automated data-quality assessment, feature discovery, knowledge extraction and reproducibility controls, and tests it on six public scientific datasets from astroparticle physics, chemistry, medicine, materials science, civil engineering and geography (569 to 21,263 records). An automated audit of the unmodified data found that 40.1% of the records in the superconductivity dataset share an identical feature vector with at least one other record, with 2,375 groups carrying conflicting critical temperatures, and that a conventional robust z-score rule flags 9.9–57.5% of rows as outliers, which makes it unsuitable as an unsupervised filter. In controlled corruption experiments, cross-validated confident learning ranked injected label errors with an area under the ROC curve (AUROC) of 0.78–0.99, and filtering the flagged samples raised test accuracy at 30% noise from 0.705 to 0.787 (QSAR biodegradation) and from 0.809 to 0.899 (breast cancer). Model-based imputation reduced the normalised reconstruction error on superconductivity from 1.00 (mean imputation) to 0.23, while native missing-value handling in gradient boosting matched imputation in predictive accuracy on the smaller datasets. Local outlier factor detected injected gross errors best (AUROC 0.82–0.91), and robust filtering restored the concrete R² from 0.19 to 0.90 at 5% corruption. Near-duplicates inflated reported R² by up to 0.036 under random splitting compared with group-aware splitting. Automated feature discovery raised the cross-validated R² of a linear model on concrete strength from 0.605 to 0.873, and a log-linear sparse regression recovered the exponents of six physical laws with a maximum error below 0.01 at 10% noise. Knowledge extraction identified the water-to-cement ratio as the first decision rule for concrete strength and thermal-conductivity descriptors as dominant for superconductivity, in agreement with domain knowledge, with importance-rank correlations of 0.95 and 0.69 across bootstrap retraining. Fixed seeds made ten repeated runs bit-for-bit identical, whereas uncontrolled runs varied by up to 0.053 in test score, mostly because of the data split. All quality checks scaled close to linearly with data size. The results show that data-quality assessment, group-aware evaluation and provenance tracking change scientific conclusions measurably and should be treated as standard parts of machine-learning-based research.