New UH Algorithm Finds Direction in 2D Data

A University of Hawaiʻi at Mānoa student-led team has published a new algorithm for finding direction in noisy two-dimensional data, and the paper appeared in AIP Advances on February 6, 2026. Led by physics undergraduate Jeffrey G. Yepez, the method compares a measured grid of numbers with a rotated reference grid using the Frobenius norm, a “distance formula” for tables of numbers, and picks the rotation with the smallest difference. It was built to locate neutrino sources such as nuclear reactors, per the University of Hawaiʻi news release of February 19, 2026.

What is the new University of Hawaiʻi algorithm?

The paper is titled “Algorithm to extract direction in 2D discrete distributions and a continuous Frobenius norm” and is published as AIP Advances 16, 025110 (2026), DOI 10.1063/5.0315079. Its authors are Yepez, Jackson D. Seligman, Max A. A. Dornfest, Brian C. Crow, Professor John G. Learned and Viacheslav A. Li. The university release says the students were guided by Learned and mentored by Li, a UH alumnus and Lawrence Livermore National Laboratory staff scientist, with funding from the Consortium for Monitoring, Technology and Verification.

How does it find direction?

The paper’s abstract describes four steps:

• Model: a reference dataset with a known direction is modeled as a 2D Gaussian distribution and binned into a histogram, which is stored as a square matrix.
• Rotate: the matrix is rotated around its center and compared with the measured dataset using the Frobenius norm of the difference.
• Fit: the resulting values are fit with an absolute sine function.
• Read the angle: the location of the minimum of that fit is the angle closest to the true direction of the measured data.

The authors also derive a continuous analog of the norm, which they call the continuous Frobenius norm of the difference. They show that a first-order approximation for two similar Gaussian distributions takes the form of an absolute sine function, and they validate the relationship with computer simulations. Their conclusion calls it “a reliable and interpretable metric” for directionality when the data is bell-shaped and close to Gaussian.

What did the simulations show?

The authors ran Python simulations with grids from 128 by 128 down to 4 by 4 bins and event counts up to one million. The university release says the method works especially well with high-resolution data and large datasets. The paper adds a nuance: at higher resolution, the fit converges more slowly as event counts grow, which the authors say suggests that low event rates could be partly offset by larger detector segments, at the cost of higher angular uncertainty. They call the best balance between resolution and event statistics an open question. The paper also states that it does not address the angular uncertainty of the result; that topic is reserved for another paper.

What has happened since?

The university release says the project began with simulated neutrino data to locate nuclear reactors and that further studies are underway. As of October 13, 2026, we found no published follow-up that applies this method outside neutrino detection. The paper’s version history shows the first arXiv posting on June 20, 2025 and a revised version on February 27, 2026.

Where else could it be used?

The university release lists astronomy, medical imaging, weather mapping and any system that relies on pattern recognition as potential uses. The paper names neutrino detection, multimodal astronomy and machine learning, and says the framework may extend to 3D scalar fields. These are stated possibilities, not demonstrated results. The paper reports no tests on medical images, no machine learning benchmarks and no cost or compute savings, so we make no such claims here.

Is there a small-business angle?

Not directly today. This is a research method for scientists and engineers who work with detector data, and no product, software release or pricing exists that a small business could buy or use. Our read: the useful lesson for owners is the one our guides on AI for small business and AI and small business already make, that most of the value in AI comes from tools that are proven and in use, so a result like this is worth watching rather than acting on. It is also a good example of students and mentors extending what people can do with data, rather than replacing anyone.

Questions about the algorithm

Is this a cryo-EM or medical imaging method?

No. Neither the university release nor the paper mentions cryo-EM. Medical imaging appears only in the release’s list of potential applications.

Does it reduce AI training costs?

Neither source makes that claim. The paper presents the method for determining direction in 2D data and reports simulations of accuracy, not training or compute costs.

Where was it published?

In AIP Advances, volume 16, article 025110, published February 6, 2026, as a featured article per the university release.