What a battery breakthrough reveals about AI and scientific memory

At Microsoft Build 2026, Microsoft Discovery showcased a closed-loop system integrating AI, machine learning, and lab feedback to design a new organic replacement candidate aimed at reducing dependence on vanadium in grid energy storage. The system combined computational predictions with wet-lab measurements to refine molecular designs, marking the first end-to-end validated scientific result with in-lab validation. Discovery Engine with CLIO mode and Bookshelf, built on GraphRAG, evolved a benzo[c]cinnoline scaffold by coordinating multiple agents to align with scientific objectives. The approach emphasized preserving both positive and negative results to guide future design cycles, addressing the challenge of aligning generative models with imprecise scientific goals.

Scientific AI traditionally focuses on generating ideas, but the new bottleneck lies in aligning these ideas with practical objectives, such as efficiently scheduling experiments and leveraging failures to shape exploration. Microsoft’s system addressed this by using Discovery Bookshelf to maintain a structured record of hypotheses, predictions, outcomes, and contradictions over time. The Bookshelf enabled the system to escape local minima by considering a global view of the task and learning from negative results, which are often overlooked in scientific publishing. This structured knowledge retention transformed campaign-specific lessons into reusable organizational assets.

The campaign, conducted in partnership with Yale Engineering, Canam Bioresearch, and Pacific Northwest National Laboratory, progressed through three rounds of molecular design. The first two rounds emphasized broad exploration, producing predominantly negative results that were captured in Discovery Bookshelf. These negative results revealed limitations in the tooling and informed the workflow adjustments for the third round, where four of six explorations succeeded. The preserved evidence allowed the system to adapt its approach based on accumulated knowledge rather than repeating prior failures.

Discovery Engine’s recalibration of trust in predictive models highlighted the system’s ability to extract value from imperfect tools. By treating redox potential predictions as relative signals rather than precise estimates, the system adjusted its workflow to prioritize ranking candidates effectively. This shift enabled the Engine to evolve the design campaign itself, moving beyond isolated candidate optimization to a more informed, iterative process guided by thematic knowledge accumulation.