Foundation models
Learning transferable representations of scientific data, and investigating how knowledge from one experiment can support another.
Vinicius Mikuni · Physicist
I'm an Associate Professor at the Kobayashi–Maskawa Institute, Nagoya University. My group develops machine learning methods to understand the physical world, from particle collisions to the cosmos.

Research directions
Learning transferable representations of scientific data, and investigating how knowledge from one experiment can support another.
Building fast, faithful simulations of particles, detectors, and physical systems using diffusion models and scientific structure.
Developing methods for unfolding, uncertainty quantification, and anomaly detection to extract more information from experimental data.
Selected work
Foundation models · 2026We distill the knowledge of OmniLearned into compact, attention-free networks for top-quark jet tagging. Message passing and pretrained teacher predictions improve background rejection, especially at the low signal efficiencies relevant for real-time triggers.
Read the paper
Foundation models · 2026We compare classification, flow matching, and masked-particle pretraining within OmniLearned. The best objective depends on the downstream task and available labels: masking complements classification when labels are scarce, while generative transfer requires generative pretraining.
Read the paper
Cross-domain learning · 2026Knowledge learned from collider data transfers to few-GeV neutrino interactions in MINERvA. Pretrained OmniLearned and ParticleViT models improve energy regression and event classification at matched compute budgets, despite large differences in detectors, energy scales, and physics processes.
Read the paper
Fast simulation · 2026A point-cloud diffusion model learns lower-multiplicity oxygen collisions before adapting to lead collisions. We assess generated high-multiplicity events through particle observables, collective flow, reconstructed jets, and classifier-based fidelity tests.
Read the paper
Fast simulation · 2026We generate gamma-ray and proton shower images for H.E.S.S. with score-based diffusion models. Tests of image structure, correlations, and gamma–hadron separation show improved fidelity over GANs, including the more challenging proton showers.
Read the paper
Anomaly detection · 2026Using OmniLearned to select events in CMS data, we test a complete background estimate. The estimate describes validation regions but does not accurately model the signal region, motivating further scrutiny of both the events and the method.
Read the paper
Foundation models · 2025Can a model trained on collider data help with cosmology? OmniCosmos transfers representations from OmniLearned to predict cosmological parameters and the velocities of halos and galaxies.
Read the paper
Foundation models · 2025An updated architecture, training on more than one billion jets, and open software bring shared representations to top tagging, ATLAS b-tagging, and anomaly detection with CMS data.
Read the paper
Generative models · 2025A diffusion architecture combining convolutional layers and a latent transformer performs super-resolution and forecasting for turbulent flows, with uncertainty estimates from sampling.
Read the paper
Foundation models · 2025OmniLearn uses joint generation and classification to build transferable jet representations, supporting tagging, simulation, likelihood-ratio estimation, and anomaly detection.
Read the paper
Generative models · 2024Point-Edge Transformers and diffusion models generate complete collider events with particle identities and kinematics, while testing event-wide conservation laws.
Read the paperBackground
My research connects high-energy physics, astrophysics, nuclear physics, and fluid dynamics through shared challenges in learning from complex data.
Before joining Nagoya University, I was a postdoctoral research fellow at Lawrence Berkeley National Laboratory's NERSC, working on anomaly detection, unfolding, and surrogate models for detector simulation.
My PhD research with the CMS experiment used machine learning to distinguish rare signals in particle collisions, reconstruct objects, and mitigate pileup.
View my CVStudents, postdocs, and researchers interested in AI for the physical sciences are welcome to get in touch. Send me an introduction and tell me about your interests.