数据和代码:PBDB-T有机太阳能电池(415条记录建模集)的接受器效率和探索性配对的机器学习预测

This archive contains the data, code and results underlying the manuscript on PBDB-T acceptor efficiency prediction. All tables and figures in the manuscript correspond to this version. Modelling set: 415 records (330 development, 85 fixed test). The curated table holds 416 records; one record (IDT-DPP) was excluded from modelling because its source DOI does not report a PBDB-T device PCE. Label verification status of the curated set: - 46 records settled against the original papers (41 labels kept, 5 raised to a higher value verified in the primary literature: ID-MeIC 6.46%, O-IDTBR 8.90%, IEICO-4F 10.25%, IT-2F 12.08%, ITDI 8.00%); - 3 records still under review (2 with a higher value in the supplementary material, 1 with an unresolved acceptor identity, ITCPTC); - 366 records not yet checked paper by paper; - 1 record (IDT-DPP) with a source that does not support its label, excluded. Contents: - curated_dataset/: 416 curated structure-label records with a boolean column model_included, the label revision log, the adjudication table, the record-level higher-value screening and the per-receptor verification ledger. - model_inputs/: primary_416.csv (the 415 modelled records), Morgan fingerprints, RDKit descriptors, the frozen 25 outer-fold indices, scaffold groups. - analysis_results/: per-fold results of the 375 model fits, locked configuration and test metrics, candidate scores, prediction intervals (molecule-level and DOI-grouped), diagnostics, control experiments. - gaussian_inputs/: seven Gaussian 16 inputs (planar start with a +/-0.008 A ripple), named according to the locked ranking. - code/: Python scripts that regenerate every table and figure. - SHA256SUMS.json: checksum and size of every file. Reported performance (fixed test set, current label version): R2 = 0.3531, Pearson r = 0.6090, MAE = 2.1248 pp, RMSE = 2.7143 pp. Known limitations: one numerical label issue (ITCPTC) still awaits acceptor-identity confirmation. The 25 outer-fold scores were also used to select among 15 representation x algorithm configurations, so they are model-comparison scores rather than unbiased generalisation estimates. The fixed test labels were inspected during earlier analysis; the results are a retrospective evaluation on a historical split, not a prospective validation.