ConGen-R v0.3.0: Source-Aware Validation, Uncertainty Calibration, and Reliability Assessment for Concrete Compressive Strength Prediction

Abstract

ConGen-R (Concrete Generalization and Reliability Tool) is bilingual research software developed to accompany the study “When High Accuracy Fails to Generalize: Source-Aware Validation and Uncertainty Calibration for Concrete Compressive Strength Prediction.” It addresses the risk that random observation-level data splitting may substantially overestimate the performance and engineering portability of machine-learning models for concrete compressive-strength prediction. The software implements publication-disjoint validation, mechanism-informed feature engineering, external dataset transfer assessment, split-conformal uncertainty calibration, publication-cluster bootstrap inference, training-support diagnostics, and material-domain warnings. HistGradientBoosting and a 300-tree Extra Trees model are provided for algorithmic comparison. The interactive application supports individual-mixture and batch CSV analysis and reports point predictions, an empirical 90% cross-publication prediction interval, inter-model disagreement, standardized nearest-neighbour distance, training-support categories, and application-risk warnings. This release contains: a reproducible Python 3.12 analysis package; command-line validation and prediction workflows; a Streamlit application; a standalone Traditional Chinese/English browser interface; trained HistGradientBoosting and Extra Trees deployment models; input auditing and physical-range checks; example input files; model-reconstruction and verification tools; automated core tests; citation, licensing, and reproducibility documentation. The primary analysis used 3,013 usable mixture–age observations from a single Zenodo benchmark containing records originating from 58 source publications. The DOI labels supplied with the benchmark were retained as publication-level grouping identifiers for leave-one-publication-out validation. The original benchmark is available as “Concrete Materials Data Extraction Benchmark, Version v1” at https://doi.org/10.5281/zenodo.22132837. Third-party row-level benchmark and UCI data are not redistributed in this software archive. Users wishing to reproduce the complete validation analysis must obtain the datasets from their original repositories and comply with the applicable licences. The browser application contains trained model parameters and training-support diagnostics required for local inference, but not the third-party row-level observations. ConGen-R is intended for research, teaching, mixture screening, and preliminary experimental planning. Its predictions, uncertainty intervals, support categories, and risk warnings are diagnostic aids rather than certified engineering acceptance criteria. The software does not replace trial batching, compressive-strength testing, mixture qualification, structural design, applicable standards, or professional engineering judgement. Predictions for oyster-shell concrete, recycled aggregates, or other materials not explicitly represented in the training benchmark should be treated as exploratory and confirmed experimentally. ConGen-R v0.3.0 is released under the MIT License.