A postdoc spent a year and $78,000 testing 520 material compositions. That covered 0.0005% of the search space. This is the trap most R&D labs are stuck in: physical screening that's Edisonian — testing carbon filaments one batch at a time, guided by intuition. The math is brutal. For a lead-free perovskite solar cell, the viable space is ~100 million compositions. At 3-5 syntheses a week, $150 each, you exhaust the budget before nearing the optimum. We rebuilt it as a closed loop. A graph neural network pre-trained on 50,000 DFT-calculated structures predicts bandgap and formation energy in milliseconds. Bayesian optimization (Expected Improvement) picks only the compositions where predicted performance is high or model uncertainty is worth resolving. Everything else never gets synthesized. The result: the top 0.1% found in 80-120 targeted experiments, not 520 random ones. Reagent cost drops from ~$78K to $12-18K. The hard part was never the algorithm. It's wiring it into instruments that all speak different proprietary dialects (Hamilton, Tecan, Agilent) — without handing your data and IP to a vendor's black box. Until you do, your scientists are just human middleware, re-entering data by hand between machines that won't talk to each other. If you run a materials or pharma lab: what's blocking you from closing the loop — the instrument integration, the cold-start training data, or GxP audit-trail compliance? #MaterialsDiscovery #LabAutomation #BayesianOptimization