High-throughput experimentation (HTE) data from R. Schlama, J. W. Sin, R. P. Burwood, K. Püntener, R. Bigler and P. Schwaller, "Swarm Intelligence for Chemical Reaction Optimisation", arXiv:2509.11798 (2025), doi:10.48550/arXiv.2509.11798; journal version "Swarm intelligence for chemical reaction optimization", Chem (2026), doi:10.1016/j.chempr.2026.103035. The dataset keeps the paper's two experimental HTE datasets together (989 reactions). Both were run in 1 mL vials in 96-position reaction blocks on a glove-box-enclosed Unchained Labs HTE platform:

SI section and figure numbers refer to the arXiv version.

(1) Pd-catalysed sulfonamide N-arylation of 1-chloro-4-methoxybenzene (10 mg, 70.1 µmol, 1.00 equiv) with benzenesulfonamide (13.8 mg, 1.25 equiv): Pd precursor (0.03 equiv), monophosphine ligand (0.03 equiv), base (2.5 equiv; or NaTFA 1.25 equiv + DBU 1.25 equiv), solvent (150 µL), 70 °C, 18 h. 585 reactions (SI section 7). The authors collected these to build a virtual benchmark on which alpha-PSO and Bayesian optimisation were compared in silico. 96 of them match a curator-reconstructed Sobol sample from the authors' initialisation (seed 9, as in the Suzuki configurations). This is an inferred match to the paper's landscape sample, not author-supplied physical plate provenance. The paper does not list these reactions or explain how the other 489 were chosen; reproducible reconstruction is provided in prov/.

(2) Pd-catalysed Suzuki-Miyaura coupling of 3-chloropicolinonitrile (72.2 µmol, 1.00 equiv) with (4,6-dimethoxypyrimidin-5-yl)boronic acid (1.25 equiv): Pd precursor (0.03 equiv), monophosphine ligand (0.03 equiv), base (2.0 equiv), solvent alone (86 reactions) or solvent plus water with SURF solvent fractions 0.75/0.25 (318 reactions; the paper gives no ratio), 70 °C, 20 h. 404 reactions (SI section 8; main-text Figure 6). The campaign started from a 96-reaction Sobol-sampled plate, followed by two optimisation iterations run in parallel by alpha-PSO and by Bayesian optimisation (qNEHVI), with 96 suggestions per method per iteration (alpha-PSO with c_local = 1.5, c_social = 0.5, c_ml = 2.0, chosen after landscape analysis of the Sobol plate; BO with qNEHVI in the Minerva workflow). Unique suggested conditions first appearing in each iteration: 96 in iteration 1, 136 in iteration 2, 172 in iteration 3. A condition suggested by both methods, or in both iterations, appears once. Each reaction's notes give suggestion iteration and method, reconstructed from the authors' files. This establishes coverage of published suggestions/results, not complete physical plate execution or the absence of discarded unpublished measurements. Physical reruns are not established. The total solvent volume is not reported for this campaign and is recorded as an unmeasured amount with the volume fractions.

Reactions were diluted with 4/1 MeCN/H2O (500 µL) and shaken 20 min, and a 50 µL aliquot was analysed by UPLC-MS (Waters ACQUITY UPLC I-Class with QDa). Outcomes are the authors' LC area-percent values. "AP yield" is recorded as an AREA measurement in percent: uncalibrated area percent, not a quantitative yield; internal-standard use is not reported and its optional boolean is left unset. "AP selectivity" is recorded as a CUSTOM SELECTIVITY measurement in percent. AP yield values of 0 are retained exactly as reported in the source. The paper gives no formula for AP selectivity but treats AP yield above AP selectivity as physically impossible (SI section 3.1); the source files nevertheless report AP selectivity 0 for every reaction with AP yield between 0.5% and 5% (105 reactions). These may reflect a reporting cut-off; this interpretation is not confirmed and the values are retained as reported.

The ligand named "P(tBu)3" in both source files is recorded as PtBu3·HBF4 in the sulfonamide reactions and as free P(t-Bu)3 in the Suzuki reactions, according to the respective HTE drawings in SI Figures 35 and 39. This drawing-based assignment does not independently establish bottle identity. The separate final-journal upscale uses the HBF4 salt; it is a different procedure and does not resolve the HTE reagent-form uncertainty. The precursor drawn as "[Pd(dba)3]" in SI Figures 36 and 40 is recorded as Pd2(dba)3, following the SURF files and main-text Figure 6a.

Source data: the authors' SURF files pd_sulfonamide_SURF.csv and pd_suzuki_SURF.csv (github.com/schwallergroup/alphaswarm, data/HTE_datasets, commit 641b258, MIT licence). The SURF files give starting materials and products as SMILES and all other components by name. The curator added structures for the Pd precursors, ligands, bases and solvents. The original curation used parallel agent resolutions, source drawings, available CAS references from the same laboratory's Minerva data, and PubChem records; retained evidence is in structure_evidence/. The correction adds runnable structure consistency checks and a documented targeted source-drawing review in analysis/. These checks are not exhaustive independent confirmation of every experimental identity. Each reaction's notes give its source file and row. Curated for the Open Reaction Database via Tokens for Progress.


Correction of TFP submission sub_536xxz341y, 2026-10-02: internal-standard uncertainty and inference language clarified; provenance reconstruction, source-transfer checks, structure checks and validation evidence included. Original curation assistance: Claude Code with parallel agents, ord-schema and RDKit. Correction and independent audit assistance: Codex with parallel agents. No wet-lab experiment or formal ORD acceptance is implied.
