🔬 The Limits of AI in Science - Why We Need Self-Driving Labs — Joseph Krause, Radical AI

The representation problem

Unlike small molecules that can be encoded as SMILES or SELFIES strings, materials like alloys involve complex variables—microstructure, processing methods (additive vs. casting), and supply chain costs—that cannot be captured in text format.

No one-shot material discovery

There is no single AI model that can predict a new material that will end up in an iPhone or Starship because performance depends on synthesis, characterization, and manufacturing variables that generative models cannot capture alone.

Experimental ground truth

Radical AI's thesis centers on self-driving labs (SDL) that combine AI hypothesis generation with automated synthesis and characterization, creating a closed feedback loop where physical experiments validate and refine predictions.

High-throughput alloy discovery

The company has synthesized 1,200 alloys in 6 months (300 novel, never-before-seen compositions), with approximately 10 showing performance exciting enough for industrial applications in extreme environments.

The fragmentation bottleneck

Traditional 15-30 year materials timelines stem from disconnected handoffs between academia (discovery), small companies (testing), and large industry (optimization), with no data sharing across the qualification pipeline.

The 10-year qualification barrier

Even promising alloys face decade-long FAA or MIL-SPEC qualification processes for aerospace applications, creating late-stage 'gotchas' similar to clinical trials; DARPA is exploring additive manufacturing methods to compress this timeline.

Concurrent engineering approach

Radical AI advocates for designing materials alongside products—adopting a SpaceX concept—rather than using 1950s-70s alloys in modern applications, to ensure new materials meet specific industrial specs from inception.