arXiv:2605.26285v1 Announce Type: new Abstract: This work addresses the problem of learning the dynamics of high-dimensional probability densities over time using unlabeled samples, without assuming access to trajectory information. We introduce two-parameter flows that learn only sampling-time transports from a base distribution to each marginal and then extract a physics-time velocity by regressing on coupled synthetic trajectories. We prove that the resulting physics-time dynamics are unique and inherit regularity from the sampling-time transports. Because we can build on standard, well-dev
The recent advancements in machine learning, particularly in generative models and physical simulations, are enabling a new approach to understanding and predicting complex systems with unlabeled data.
This research introduces a novel method for learning the dynamics of complex physical systems from raw, unlabeled data, significantly reducing the reliance on meticulously curated datasets or explicit trajectory information.
The ability to infer underlying physical dynamics from high-dimensional probability densities without direct trajectory data opens new avenues for AI application in scientific discovery, complex system modeling, and potentially even control.
- · AI researchers and data scientists
- · Scientific research institutions
- · Industries relying on complex simulations (e.g., materials science, climate mode
- · Traditional simulation methods requiring explicit data acquisition
Improved models for predicting the behavior of unobserved or poorly understood physical systems become feasible.
This could accelerate the design and optimization of new materials, drugs, or industrial processes by simulating their dynamics more effectively.
A deeper AI-driven understanding of fundamental physical laws could emerge, potentially leading to paradigm shifts in various scientific disciplines.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG