Periodic Labs
An AI for Science startup focused on using AI foundation models to accelerate materials discovery, extending from semiconductors to the screening and synthesis planning of superconductors.
Timeline (last 90 days)
Periodic Labs' core claim is: materials discovery can be modeled as a search problem, and search efficiency depends on the accuracy and coverage of the prediction model; when the model can simultaneously predict structural stability, synthetic feasibility, and target properties, the hit rate of experimental validation will increase substantially.
Cubuk believes that representation learning of crystal structures is general, but property prediction requires fine-tuning for different physical mechanisms, so Periodic Labs needs to build a multi-task framework rather than a single-purpose predictor.
Periodic Labs adopts the Forward Deployed Engineering model, in which engineers are directly embedded in customer or experimental teams and quickly adjust models based on actual feedback; this term borrows a concept from Palantir.
Periodic Labs needs to build its own automated experimental platform to let models learn quickly from failures through a closed loop of synthesis and characterization, because public data is insufficient to support learning for inorganic retrosynthesis tasks.
Periodic Labs has unveiled Neon, a materials science model trained using 1,300 H200 GPUs and months of experimental data, focusing on superconductors, magnets, and semiconductor materials.
All Facts (5)
Periodic Labs' core claim is: materials discovery can be modeled as a search problem, and search efficiency depends on the accuracy and coverage of the prediction model; when the model can simultaneously predict structural stability, synthetic feasibility, and target properties, the hit rate of experimental validation will increase substantially.
50%UnverifiedCubuk believes that representation learning of crystal structures is general, but property prediction requires fine-tuning for different physical mechanisms, so Periodic Labs needs to build a multi-task framework rather than a single-purpose predictor.
50%UnverifiedPeriodic Labs adopts the Forward Deployed Engineering model, in which engineers are directly embedded in customer or experimental teams and quickly adjust models based on actual feedback; this term borrows a concept from Palantir.
50%UnverifiedPeriodic Labs needs to build its own automated experimental platform to let models learn quickly from failures through a closed loop of synthesis and characterization, because public data is insufficient to support learning for inorganic retrosynthesis tasks.
50%UnverifiedPeriodic Labs has unveiled Neon, a materials science model trained using 1,300 H200 GPUs and months of experimental data, focusing on superconductors, magnets, and semiconductor materials.
50%