Yield Prediction and Condition Recommendation on High-Throughput Experimentation Datasets: Encoding Effects and Noise Ceilings

High-throughput experimentation (HTE) supplies enough reaction data to train machine-learning models, yet yield-prediction models are typically reported on a single dataset, and both their cross-dataset stability and their attainable ceiling are rarely examined. Using a single pipeline without per-dataset tuning, this work compares five model architectures under three categorical encodings on thirteen HTE datasets drawn from five public source collections, and estimates a noise ceiling for every dataset that repeats a condition. Accuracy is governed by the encoding rather than by the algorithm: switching a multilayer perceptron from integer to one-hot encoding raises its mean held-out R2 by 0.168. Across datasets, the spread in validation-selected held-out R2 tracks the noise ceiling, with a mean capture ratio of 86.35%. Attribution rankings show some regularity: the reactant slots rank at the top, followed by temperature and the additive, while the equivalence and metal slots carry the smallest mean shares. The limits are equally clear: substrate-disjoint and ordered cross-validation both give negative R2, and no cross-dataset transfer succeeds, so the framework supports interpolation only. On the only dataset with enough distinct reactant pairs, predicting unseen combinations of known reactants reaches R2 = 0.58, whereas predicting unseen reactants with molecular fingerprints reaches 0.41. Ranking by predicted yield outperforms expected improvement, whose uncertainty signal neither ranks the error nor is calibrated.