Joseph Krause, co-founder and CEO of Radical AI, outlined the key roadblocks holding back AI in scientific discovery - and why automated "self-driving labs" are the only way to break through them. Speaking on the Latent Space podcast, Krause said current AI models simply cannot handle the unstructured, multi-modal data that dominates fields like materials science.
In areas such as AI for biology or materials screening, data is often neatly structured. But materials science demands much more. To properly predict a new material, an AI system must absorb not just molecular structures but also supply chain constraints, manufacturing cost, microstructure, and processing routes like additive manufacturing versus casting. This wide range of data - sometimes encoded as "smile strings" in computational chemistry - overwhelms models that were designed for narrower problems.
These multi-modal data hurdles require more than general AI knowledge; specialized AI learning path for research scientists covering lab automation and data modeling can help bridge the gap between theory and experimental reality.
Why current AI models fall short
Krause offered a blunt assessment: "There is no one model that can one-shot a new material that ends up in your iPhone that ends up on Starship." The issue isn't just identifying a promising compound - it's understanding whether it can be manufactured at scale and integrated into real products. Current models, he said, lack the integrated reasoning needed to span discovery, manufacturability, and commercial viability in a single workflow.
This fragmentation means scientists still rely on manual cycles of experiment design, testing, and data interpretation. AI can assist with parts of the process, but it can't yet own the entire discovery loop.
How self-driving labs close the loop
The solution, Krause argued, lies in self-driving labs - closed-loop systems where AI designs experiments, robotic systems execute them, and the results feed directly back into the AI for the next round of decisions. This autonomous cycle eliminates the human bottleneck of analyzing data and deciding what to test next, compressing what might take months into days or hours.
In Krause's vision, these labs don't stop at suggesting materials; they also gather the real-world performance data needed to move a candidate from the bench to a product. When the system tests a formulation, it captures processing conditions, yields, and failure modes - all of which sharpen the model's ability to predict success in manufacturing.
Experimental data as the foundation
Krause stressed that the race to automate discovery hinges on one thing: "What makes us different is our deep belief in experimental data." Theory and simulation can point in the right direction, but without measured results from physical experiments, AI predictions remain unvalidated. He noted that many organizations are now investing in self-driving labs that can generate and learn from that experimental data across the full discovery-to-development pipeline.
However, too often that data remains siloed, trapped in the notebooks of individual scientists or incompatible formats from different equipment. Making it machine-readable and feeding it back into the AI loop is a technical challenge that must be solved for autonomous labs to deliver on their promise.
Why this matters for science and research professionals
For researchers and lab managers, the path forward is clear: AI-powered automation won't just accelerate individual experiments - it will reshape how labs are staffed, funded, and measured. Professionals who can bridge the gap between AI, lab automation, and domain science will be the ones driving the next generation of materials and drug discovery. Keeping current on these shifts through resources like AI for Science & Research can help teams spot practical use cases and avoid investing in tools that don't integrate with real experimental workflows.
Your membership also unlocks:
