New AI cracks complex engineering problems faster than supercomputers

The new artificial intelligence framework is a general approach that can quickly predict solutions to common and time-consuming mathematical equations that require creating models of how fluids or electrical currents propagate through different geometries, such as those involved in standard engineering tests.

Details about the study appear in natural computational science.

The framework, called DIMON (Diffeomorphic Mapping Operator Learning), solves mathematical problems called partial differential equations that are ubiquitous in almost all scientific and engineering research. Using these equations, researchers can transform real-world systems or processes into mathematical representations of how objects or environments change over time and space.

“While the motivation for developing it came from our own work, we believe this solution will have a huge impact in various engineering fields in general because it is so general and scalable,” said Johns Hopkins Biomedical Engineering, who co-led the project. Research Professor of Medicine. “It can solve basically any problem in any field of science or engineering, solving partial differential equations for multiple geometries, such as crash testing, orthopedic research, or other complex problems where shape, force, and materials change.”

In addition to demonstrating DIMON’s applicability to solving other engineering problems, Trayanova’s team tested the new AI on more than 1,000 cardiac “digital twins” – highly detailed computer models of real patient hearts. The platform is able to predict how electrical signals propagate through each unique heart shape, allowing for higher prognostic accuracy.

Trayanova’s team relies on solving partial differential equations to study arrhythmias, an abnormal behavior of the heart’s electrical impulses that cause irregular beats. Using cardiac digital twins, researchers can diagnose whether patients are likely to suffer from this often fatal disease and recommend treatments.

“We are bringing new technologies into the clinic, but many of our solutions are very slow. It takes us about a week to go from scanning a patient’s heart and solving partial differential equations to predicting whether a patient is at high risk for a sudden heart attack. cardiac death and what is the best treatment plan,” said Trayanova, director of the Johns Hopkins Consortium for Innovation in Cardiovascular Diagnostics and Therapeutics. “With this new artificial intelligence approach, we are getting solutions incredibly quickly. The time to calculate a heart digital twin prediction will be reduced from hours to 30 seconds, and will be done on a desktop computer rather than on a on supercomputers, allowing us to make it part of daily clinical workflow.

Partial differential equations are usually solved by breaking down complex shapes, such as airplane wings or body parts, into grids or grids of small elements. Then work on each simple part and regroup. However, if these shapes change (such as collisions or deformations), the mesh must be updated and the solution recalculated, which can result in slow and expensive computations.

DIMON solves this problem by using artificial intelligence to understand how physical systems behave on different shapes without having to recalculate everything from scratch for each new shape. Instead of dividing shapes into meshes and solving equations over and over, AI can predict how factors like heat, pressure, or motion will behave based on the patterns it learns, allowing it to do things like optimizing designs or modeling shapes. Faster and more efficient on the job – specific scenarios.

The team is integrating cardiac pathologies that lead to arrhythmias into the DIMON framework. Minglang Yin, a postdoctoral fellow in biomedical engineering at Johns Hopkins University who developed the platform, said that because of its versatility, the technology could be applied to shape optimization and many other engineering tasks that require iteratively solving partial differential equations for new shapes.

“For each problem, DIMON first solves a partial differential equation on a single shape and then maps the solution to multiple new shapes. This shape transformation capability highlights its tremendous versatility,” Yin said. “We are very excited to be able to use this to solve many problems and make it available to the wider community to accelerate their engineering design solutions.”

Other authors include Nicolas Charon of the University of Houston, Ryan Brody and Mauro Maggioni (co-leads) of Johns Hopkins University, and Lu Lu of Yale University.

This work was supported by NIH grants R01HL166759 and R01HL174440; a grant from the Leducq Foundation; the Heart Rhythm Society Fellowship; U.S. Department of Energy grants DE-SC0025592 and DE-SC0025593; NSF grants DMS-2347833, DMS-1945224, and DMS-2436738; and Air Force Research Laboratory awards FA9550-20-1-0288, FA9550-21-1-0317, and FA9550-23-1-0445.