Three threads of our work from the past two years are now mature enough to wire together into one autonomous discovery loop for the data-scarce, unbenchmarked regime of materials and device discovery.

1. Hypothesis generation: DToR. A local-first deep-research agent with a tree-structured orchestrator that adaptively expands and prunes research branches. Across 27 nanomaterials/device topics in a 44-agent benchmark, its reports achieved a ~79% mean pairwise win rate against commercial deep-research systems, running entirely on consumer-level hardware with open-source LLMs (arXiv).

2. Device digital twins: T³. Text-Twin-Translation converts unstructured literature into structured knowledge (TextGrad-optimized extraction, +21.8% BLEU), trains a device-topology-aware GNN as a digital twin (92.3% sensitivity prediction accuracy), and screens candidates at scale: 123.2 million PubChem compounds (OpenReview, code).

3. Rapid physical validation: RAPIDS. An atomistic validation engine benchmarking MLIPs against DFT over 5,567 probe–target dimers and 18 tasks, packaged as a callable tool for autonomous agents (OpenReview).

Each of these is a paper on its own. Chained together under BRAINIAC, they start to look like an AI co-scientist: DToR proposes, T³ screens, RAPIDS verifies, and the wet lab only gets involved once a candidate has survived all three gates.

The first full end-to-end test is now funded: our 2026 Schmidt Seed Fund project on PFAS-selective membrane interfaces will run this loop from autonomous design all the way to experimental validation.