Periodic trains Neon on its own lab data: 55.3% success on FrontierXRD at lower cost than frontier models
Synopsis
Periodic trained Periodic Neon on X-ray diffraction data from its own labs, reaching a 55.3% success rate on FrontierXRD, its hardest internal evaluation set, against an initial 2.7% for Kimi K2.6, the open-weight model it post-trained. The evaluation covers 134 samples from its labs that human experts take hours to resolve because of their complexity and ambiguity; Neon surpasses frontier models at a lower cost per analysis and is now deployed in its labs analyzing experiments in the search for better superconductors and magnets.
Interpretation
Periodic turns powder X-ray diffraction data produced in its labs into training signal, yielding a model, Periodic Neon, that carries out multiphase XRD analysis on its own and is already deployed in its labs' search for superconductors and magnets. Such analysis previously depended on a scientist first narrowing the plausible elements and phases inside existing software, then judging from synthesis context whether a fit is chemically sensible; as autonomous labs scale, manual analysis can no longer keep up with experiment output. On FrontierXRD, an internal evaluation set of 134 complex samples from its labs, Neon reaches a 55.3% success rate, while Kimi K2.6, the open-weight model it post-trained, starts at 2.7%.
Neon's analysis ability comes from midtraining and reinforcement learning on lab data, not from swapping in a stronger base model alone. Success in XRD analysis cannot be cheaply verified from pattern fitting alone; it requires expert judgment on whether each phase is supported by the pattern and makes chemical sense, so ordinary verifiable rewards do not apply. Periodic compared an ensemble of two agentic LLM judges with labels from three PhD experts in materials-related fields across thousands of patterns: human experts agreed with each other 77.2% of the time, the judge ensemble agreed with human experts 74.6% of the time, and with an expert consensus 84% of the time.
The accompanying scientific tooling contributes substantial analysis success on its own, beyond the model weights. The usual setup gives a model Claude Code plus open-source databases and open-source analysis software, representing a typical scientist's toolkit without internal infrastructure. With the same underlying model, Claude Opus 5, Periodic's scientific harness achieves a 3.8x higher XRD analysis success rate on FrontierXRD than Claude Code with open-source databases and software, at a similar cost per analysis.
Perspective
The demonstrated scope is powder XRD phase and proportion analysis: the 134 FrontierXRD samples come from its labs, and generalization is also evaluated on 198 experimental measurements from held-out chemical systems. The intended reader is a researcher interested in materials-discovery automation and scientific agents; the material states that training is now extending to a wider set of scientific workflows across its labs.
FrontierXRD is an internal evaluation set, and success is decided by an LLM judge ensemble whose agreement with human experts is lower than agreement among human experts. The held-out chemical-system evaluation is easier than FrontierXRD, and the authors note that training compute is far below what frontier models report.