Roboticized AI-assisted microfluidic photocatalytic synthesis and screening up to 10,000 reactions per day

Outcome

I submitted sub_xp1gilyg91. Verification run vrn_v21t3b5tu3 is waiting for the curation review. It includes two Artifacts:

• dataset = art_jcues7zrbw (lu2024_photocatalysis.pb.gz, 12,197 reactions, DOI on every reaction, no IDs).
• sources = art_meqwk44teh (README, build_dataset.py, structures.py, source xlsx, Table S3 rows, validation.txt).

The ord-schema 0.8.3 validator exits 0 with 0 errors and 2 warnings. Both warnings are substrate recoveries above 100 % (106 % and 104 %) in dark controls, and those are the values the SI reports. Rebuilding from the zip gives an identical dataset.

Approach

• Source: Photocatalysis_data.xlsx, all 12,000 rows, each used once. I skipped the overlapping dataset.xlsx.
• Chemistry: the Ir-photosensitised intramolecular [2+2] of bis-allylic ethers to 3-oxabicyclo[3.2.0]heptanes, run in DMSO in an LCW fused-silica capillary under 4 × 450 nm lasers at 25 °C. Structures come from Fig. 2, SI §1/§7 and Table S1.
• Extra reactions: I added the 197 GC/NMR validation runs from SI Tables S3–S7 (yield, d.r., recovery).
• Absorbance outcomes: these are CUSTOM product measurements (stable, non_stable) with the monitoring wavelength. They are not converted to yield, because the Fig. S9 calibrations are not reported numerically.
• Flow encoding: input flow_rate, laser power in the illumination details, and reaction_time = residence time = 660/flow s.

Key findings and decisions

• S-3 structure conflict: the SI name, Table S1 SMILES and the Fig. 2 R² = H don't match the S-3 ¹H NMR, the P-3 name or the P-3 drawing. I used (E)-2-methylcinnamyl prenyl ether, which fits the NMR and the product. A reviewer may want to check this.
• P-5 name: the SI names P-5 as the 4-chlorophenyl compound, which is a copy error; it is the 4-methoxyphenyl product.
• Tables S4–S7 conditions: these tables list only flow rate and light. I assumed Ir-1, 1 mol% and 0.01 M, which are the SI §7 general conditions; entry 12 of each table matches them.
• Input amounts are for the 5 mL premix made for each set of 30 flow/laser conditions.
• Not included: the batch-vs-flow comparison and the high-concentration runs (Fig. 5D/E). They exist only as plots.
• Curator email: the validator requires record_created.person.email. I used the placeholder ord-curator@example.org rather than a personal address; whoever submits to ORD should replace it.

Tooling issues

• The pinned ord-curation Skill could not be read. This client's MCP resources/read rejected TfP's response (schema error on ttlMs/cacheScope), so I followed general ORD conventions instead of the Skill. A reviewer should check the dataset against the Skill.
• The Claude Code Remote get_session tool was not available, so report_agent_usage was not called.
• The github.com API was blocked by the session proxy, but raw.githubusercontent.com worked.
• I did not open the ORD wishlist card's live status. The card snapshot (2026-09-25) says Available.

Advice for next

If a reviewer wants yields for the screen, the absorbance→yield fits would need the authors' Fig. S9 coefficients or a refit against Tables S3–S7, using the absorbance at matching conditions.

Ran build_dataset.py (python, ord-schema 0.8.3, pandas, RDKit) locally → lu2024_photocatalysis.pb.gz (~0.5 MB): 12,000 screening reactions + 145 (Table S3) + 52 (Tables S4–S7) GC/NMR runs. Re-running from the extracted sources zip gives a byte-identical (uncompressed) dataset.

Encoding choices:

• One ReactionInput per reaction = the 5 mL premixed DMSO solution (substrate µmol, Ir µmol, DMSO 5 mL incl. solutes), syringe pump, flow_rate µL/min.
• Illumination CUSTOM (no LASER enum) with 450 nm peak and coupled power in details; flow PLUG_FLOW_REACTOR, fused-silica tubing 0.53 mm; 25 °C jacket; air.
• reaction_time = residence time 660/flow s (3.3 s at 200 µL/min per SI).
• Screening outcomes: two CUSTOM product measurements on the desired product (stable, non_stable absorbance change, float_value, wavelength 285.4/293.3/310.7 nm), analysis UV. Negative absorbances (642 non_stable, 6 stable) kept as reported.
• GC runs: YIELD % (internal standard), SELECTIVITY DR (float, or string ">10:1"), substrate recovery as AMOUNT % on a non-desired REACTANT entry; dark controls with illumination DARK and recovery only.
• Products: cis ring fusion only (diastereomer mixtures); checked by RDKit embedding and formula equality.

Validator exit 0, 0 errors; full validate_message pass: 2 warnings (recovery 106 % / 104 % in S-3/S-4 dark controls, as reported). The validator requires record_created.person.email; I used the reserved placeholder ord-curator@example.org rather than a personal address — the submitter should replace it at ORD submission.

Sources read outside TfP: article HTML (nature.com), Fig. 2 image, SI PDF (MOESM1, pdftotext).

• Reaction: Ir-photosensitized intramolecular [2+2] cycloaddition of bis-allylic ethers to 3-oxabicyclo[3.2.0]heptanes (P-1..P-5), DMSO, 4 × 450 nm lasers coupled into a liquid-core-waveguide fused-silica capillary (530 µm i.d.), 25 ± 2 °C, under air. Stocks: 0.2 M substrate, 0.0035 M Ir in DMSO; 5 mL mixture per set of 30 flow/laser conditions. Residence time 3.3 s at 200 µL/min (13.2 s at 50) → ~11 µL irradiated volume; t = 660/flow s.
• Ir-1 Ir(dF(CF3)ppy)2(dtbbpy)PF6, Ir-2 Ir(dFMeppy)2(dtbbpy)PF6, Ir-3 Ir(dF(CF3)ppy)2(bpy)PF6, Ir-4 Ir(dFMeppy)2(bpy)PF6, Ir-5 fac-Ir(ppy)3 (SI §1, Table S1, Fig. 2).
• S-3 discrepancy: SI name "(E)-(4-((3-methylbut-2-en-1-yl)oxy)but-2-en-2-yl)benzene" and Table S1 SMILES put the Me on the Ph-bearing carbon, and Fig. 2 says R2 = H, but the S-3 ¹H NMR (vinyl 6.51 s, OCH2 4.02 s, CH3 1.90 s), the product name 1,6,6-trimethyl-7-phenyl-3-oxabicyclo[3.2.0]heptane and the P-3 drawing (Me on bridgehead) all fit (E)-2-methylcinnamyl prenyl ether. Curating S-3 as CC(C)=CCOC/C(C)=C/c1ccccc1 and P-3 accordingly; documented in README.
• SI names P-5 as 4-chlorophenyl (copy error); it is the 4-methoxyphenyl product.
• Beyond the 12,000 UV screen, the SI reports 197 GC/NMR-quantified runs (Table S3: 145 S-1 conditions; Tables S4–S7: 13 each for S-2..S-5, Ir-1 1 mol% 0.01 M, incl. dark controls) with yield, d.r. and substrate recovery. Plan: include them as separate reactions in the same Dataset. Batch-vs-flow comparison (Fig. 5D/E, S10) exists only as plots → excluded.
• Absorbance→yield linear fits (Fig. S9) are not given numerically, so the 12,000 outcomes stay as absorbance (CUSTOM measurements), not yields.

• Read card via its signed download_url (hash verified). The pinned ord-curation Skill could NOT be read: this client's MCP resources/read rejects TfP's result (schema error on ttlMs/cacheScope), and directory listing is disabled. Proceeding with ORD's standard conventions (ord-schema 0.8.3 validator, provenance DOI on every reaction, no IDs).
• Claude Code Remote get_session tool is not present in this session, so report_agent_usage could not be called.
• Source: chose AI-assisted cross-species prediction/Data/Photocatalysis_data.xlsx (fetched from raw.githubusercontent.com; github.com API path blocked by session proxy). One sheet, 12000 x 10: substrate S-1..S-5, Ir-1..Ir-5, wavelength 285.4–310.7 nm, conc 0.005–0.02 M, PC 0.5–2 mol%, flow 200–600 µL/min, laser 772–4642 mW, non_stable / stable absorbance. The 10.8 MB dataset.xlsx is not used (same 12000 experiments + time traces) to avoid duplicates.
• Next: read the Nature article (fetched HTML, 200 OK) for S-1..S-5 / Ir-1..Ir-5 structures, the reaction, solvent and setup.