Lifeboat Foundation

While machine learning methods can be used for accurate flow prediction in complex environments, such as for urban structures³⁰ or turbulent fields³¹, generalizing these approaches to domains of arbitrary size and complexity remains a challenging problem. One reason is that flows near and around obstacles depend on factors associated with the fluid (i.e., Reynolds number) or domain (i.e., boundary conditions), and fixing either of these conditions puts bounds on the validity of the estimated fields. Thus, if we seek broad applicability, then we should seek the fewest set of model restrictions that together provide the most accurate flow predictions. To this end, our approach has been to deconstruct certain types of domains into individual obstacles that each maintain some level of geometrical similarity, so that a single neural network model can be used to predict flows near all structural boundaries of the domain. Flows between these structural surfaces, at a scale on the order of the obstacle diameter, are predicted using a second neural network model in series with the first. Together, this serial-modeling approach allows for rapid prediction of flows in domains that can be represented by a disjoint set of structural elements. This type of domain is common, for example, in urban and periurban areas, wherein buildings conform to a common structural motif that affects ground-level velocity fields.

Another relevant length scale is the grid size used to digitize individual domains for read-in by the model. Thus, we investigated how flow patterns can be affected when this input resolution is varied. Although our choice of grid size is somewhat arbitrary, it is dense enough to capture variation in the relevant velocity fields near individual obstacles, but not so dense that producing a large enough cohort of CFD-generated training datasets becomes computationally intractable.

Our approach can also be trained to predict flows with a variable inlet velocity, which, in the case of urban wind flow prediction, permits model parameterization in terms of current meteorological conditions. In the specific case of aerial dispersion of chemicals throughout an urban environment, our predicted flows are considered as the advective field of a drift-diffusion model of molecular dispersion. This advection field plays a central role because concentration fluctuations decorrelate in relationship with the velocity fluctuations of the advection field, and spatial heterogeneity in the flow patterns is determined by the sequence of obstacles in the flow path.

The company says the updated version responds to your emotions and tone of voice and allows you to interrupt it midsentence.

The team spotted the record-breaking triple star system because of strobing starlight caused by the stars crossing in front of each other, as seen from our position on Earth.

The team turned to machine learning to analyze vast amounts of data from TESS to spot a pattern indicating these eclipses. They then called upon the aid of citizen scientists to further filter this data to spot interesting signals.

“We’re mainly looking for signatures of compact multi-star systems, unusual pulsating stars in binary systems, and weird objects,” Rappaport said. “It’s exciting to identify a system like this because they’re rarely found, but they may be more common than current tallies suggest.”

The state-owned telecoms operator did not reveal what chips it used for its 1 trillion-parameter model, but it has a partnership with Huawei.

Fun experiment, but yeah, don’t pipe an LLM raw into /bin/bash.

Molecular dynamics (MD) is a popular method for studying molecular systems and microscopic processes at the atomic level. However, MD simulations can be quite computationally expensive due to the intricate temporal and spatial resolutions needed. Due to the computing load, much research has been done on alternate techniques that can speed up simulation without sacrificing accuracy. Creating surrogate models based on deep learning is one such strategy that can effectively replace conventional MD simulations.

In recent research, a team of MIT researchers introduced the use of generative modeling to simulate molecular motions. This framework eliminates the need to compute the molecular forces at each step by using machine learning models that are trained on data obtained by MD simulations to provide believable molecular paths. These generative models can function as adaptable multi-task surrogate models, able to carry out multiple crucial tasks for which MD simulations are generally employed.

These generative models can be trained for a variety of tasks by carefully choosing and conditioning on specific frames of a molecule trajectory. These tasks include the following.

Do #AI companies work?

“The market needs to be irrational for you to stay solvent.”

Workplace AI tools can do tasks by themselves. Getting them to stop is the problem.

The Hubbard model is a studied model in condensed matter theory and a formidable quantum problem. A team of physicists used deep learning to condense this problem, which previously required 100,000 equations, into just four equations without sacrificing accuracy. The study, titled “Deep Learning the Functional Renormalization Group,” was published on September 21 in Physical Review Letters.

Dominique Di Sante is the lead author of this study. Since 2021, he holds the position of Assistant Professor (tenure track) at the Department of Physics and Astronomy, University of Bologna. At the same time, he is a Visiting Professor at the Center for Computational Quantum Physics (CCQ) at the Flatiron Institute, New York, as part of a Marie Sklodowska-Curie Actions (MSCA) grant that encourages, among other things, the mobility of researchers.

He and colleagues at the Flatiron Institute and other international researchers conducted the study, which has the potential to revolutionize the way scientists study systems containing many interacting electrons. In addition, if they can adapt the method to other problems, the approach could help design materials with desirable properties, such as superconductivity, or contribute to clean energy production.