Thursday, February 19, 2026

Media Contact:
Desa James | Communications Coordinator | 405-744-2669 | desa.james@okstate.edu

Dr. Zeyuan Song, a recent Ph.D. graduate of the School of Chemical Engineering at
Oklahoma State University, and Dr. Zheyu Jiang, assistant professor for CHE, have achieved
a milestone rarely seen in Oklahoma’s research landscape: acceptance into the International
Conference on Learning Representations 2026, one of the world’s most competitive and
influential academic conferences in artificial intelligence and machine learning.  

ICLR ranks among the top AI venues globally – second in the field by h-index – and
is known for debuting many of the breakthroughs that have shaped modern AI, including
the variational autoencoder and the graph attention network. Each submission undergoes
a monthslong, double-blind review and rebuttal process, making acceptance highly selective.  

“I am proud of the research excellence Zeyuan achieved during his Ph.D. study in my research
lab,” Jinag said. “I have been impressed by his ability to bring in new ideas from
diverse fields in mathematics, engineering, and AI. This, when combined with a deep
understanding of the cutting-edge breakthroughs in the field, leads to this outstanding
work published in ICLR.”  

Song’s paper, titled Adaptive Fourier Mamba Operators, introduces a powerful new machine
learning framework for modeling complex natural and engineering phenomena described
by partial differential equations.   

“Imagine you are baking a cake,” Jiang said. “The temperature of the cake isn’t determined by
time alone. The outside heats faster than the inside, and the top browns more quickly
than the bottom. Partial differential equations describe changes that happen simultaneously
in space and time, like how heat moves through a cake as it bakes.”  

These types of equations govern real-world phenomena such as fluid dynamics, heat
transfer, quantum mechanics  and even the financial market.  

Unlike traditional numerical solvers, which can become extremely time-consuming to solve,
Song’s AFMO method uses a mathematically grounded neural operator framework to learn
how these systems behave, often with greater efficiency and generalizability.  

According to the paper, AFMO integrates two computational frameworks, Adaptive Fourier
decomposition, a novel signal processing technique that builds orthogonal spectral
bases tailored to the problem, and state-space models, an emerging neural network
architecture that can efficiently handle long-range dependencies, to solve general nonlinear partial
differential equations.  

“Imagine you are playing piano,” Jiang said. “Standard Fourier neural operator plays every
song on a standard piano. The piano keys are fixed, and you play by mixing those fixed
notes. It works great when the song fits that instrument well, but it can struggle
if the ‘song’ has unusual rhythms. Adaptive Fourier decomposition, on the other hand,
is like a custom keyboard tailored to the particular song one wants to play.

“Meanwhile, a state-space model is like a super-fast musician who reads the music
left-to-right and keeps a small memory of what happened so far, so they can play very
long songs efficiently. Therefore, AFMO builds a custom instrument for each song first,
and then has the super-fast musician to play it, so it has the right instrument and
efficient playing.”  

By uniting these in a novel way, AFMO can solve PDEs on irregular shapes and complex
geometries, capture sharp features and singularities, and produce results that are
both highly accurate and computationally efficient. 

“These are especially challenging problems to solve due to the intricacies of the systems
involved,” Jiang said. “They require us to think out of the box and develop truly innovative solutions.” 

In extensive testing, the method consistently outperformed leading neural operator
models across diverse benchmark problems, ranging from modeling fluid flow in airfoils
and pipes to predicting European option prices in financial mathematics.   

Song’s accomplishment represents more than an individual’s success.  

This publication is the first ICLR paper from the state of Oklahoma. Notably, this work comes from
a chemical engineering department, rather than a traditional computer science or electrical
engineering program.  

“As AI continuously transforms the world, we are in an exciting era for interdisciplinary
research,” Jiang said. “We are thrilled to see the broader impacts and implications
of this work in helping OSU recruit talented students, forming cross-department collaborations,
and competing for more federal and industry funding to support AI for Science research that pushes
toward AI capacity and workforce development in Oklahoma.”