Multi-Agent Scheduling of Instrument-Time for High-Throughput Palladium-Catalysed Cross-Coupling
Year: 2026
Type: preprint
Abstract: NMR time and a shared liquid-handling robot are scarce, noisy and only partially observable, so the practical question is which experiment to run next on which instrument. Existing reaction-optimisation benchmarks measure budgets in experiments rather than instrumenthours, and they learn chemical feasibility only by paying for infeasible runs. We present CHASE, a centralised-training decentralised-execution scheduler with three coupled modules: a belief encoder (PEB) infers per-instrument throughput and channel health from local event histories and exchanges only low-bandwidth summaries; a chemistry feasibility gate (CFG) masks chemically impossible actions with a four-layer compatibility graph; and a counterfactual attribution module (CFA) converts delayed campaign reward into per-step credit through leave-one-experiment-out surrogate scoring. On the Buchwald–Hartwig (3,955 reactions) and Suzuki–Miyaura (5,760 nanomole-scale conditions) HTE datasets, the ten Olympus benchmarks and a plant-wide Tennessee Eastman process, CHASE reaches 94.1% best-found yield under a shared 60 instrument-hour budget, +3.3 points over the strongest same-setting baseline, attains 90% of the attainable optimum with 25.6% fewer experiments, and cuts infeasible actions to 0.7–1.6% across plate, flow and fault-injection settings.
Source: ChemRxiv
Institution: Kunming University of Science and Technology
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7223178190