Learning and Optimization Platform

LOOP — the Learning and Optimization Platform — is the connective tissue of a materials-discovery pipeline, built so the metadata most critical to prediction never dies in a lab notebook or an unshared spreadsheet again. It serves two audiences at once: experimentalists, who need a frictionless place to record what they made and measured, and ML researchers, who need that record as clean, vectorizable signal for predictive models.

The platform closes a loop. ML researchers mine literature and run DFT computations to propose theoretical materials; experimentalists synthesize and characterize those candidates, then upload their results; that experimental data flows back to inform the next generation of computational models. Literature — pulled from anywhere — continually fuels new ideas and new computations, so the computational–experimental cycle sustains itself rather than stalling at any single handoff.

Before LOOP, the metadata essential to materials prediction lived in logbooks and hard-to-reach Excel files; there was no central store, and every disconnected record was signal the models never saw. LOOP parses experimental work into discretized steps, turning loose process notes into structured records an ML pipeline can read directly.

The stack is deliberately pragmatic. Django keeps the platform easy to deploy and operate; MongoDB stores the data as documents with the dense relations that map naturally onto how models consume it. On top, sentence embeddings indexed for Atlas vector search power semantic and related-material lookup — finding experiments and materials by meaning rather than exact-match metadata. LLM-capable queries and batch uploading let experimentalists search and contribute without learning the system’s internals, lowering the cost of capturing data and, in turn, surfacing even more of it.

Provenance is enforced by content hashing, with optional literature-reference linking so an experimentalist can tie a result back to the paper that inspired it — any record traces cleanly to its origin.

LOOP sits above CHAOS, the computational database that serves the DFT calculations LOOP draws on: CHAOS is the engine room for ML researchers, LOOP the shared surface for everyone. The system is live and in production under active development, today spanning 300 literature records, 60 experiments, and 200,000 computations.

Access LOOP at: s4e.ai/loop