CMU Simulates Axonal Transport 10x Faster on Bridges-2

Researchers at Carnegie Mellon University utilized the Pittsburgh Supercomputing Center's Bridges-2 supercomputer to simulate molecular transport inside branching nerve cells, according to psc.edu psc.edu. The project aims to accelerate the study of neurodegenerative conditions such as amyotrophic lateral sclerosis (ALS), where disruptions in nutrient and signal movement along axons break communication between nerves and muscles psc.edu. By replacing resource-intensive wet-lab trials with in silico models, scientists can evaluate hypotheses regarding transport bottlenecks before conducting physical experiments psc.edu.

Detailed physical models tracking fluid and nutrient dynamics across complex three-dimensional neurite branches typically create prohibitive computational demands psc.edu. To overcome these scaling constraints, graduate student Tsung Yeh Hsieh and Professor Jessica Zhang designed a surrogate framework named Graph-Autoencoder-based Latent Dynamics Surrogate (GALDS) psc.edu. The method mathematically simplifies axons and branching neurites into networks of branching pipes, compressing spatial data into primary features to compute traffic behavior efficiently psc.edu.

Computational Performance on Bridges-2

The team trained and executed the GALDS architecture using hardware allocations on the Bridges-2 system, which features 47 graphics processing unit nodes with 384 accelerators and 504 central processing unit nodes offering more than 64,000 cores psc.edu. The traffic simulations periodically utilized roughly 1,000 CPU cores, relying on node memory configurations of 256 to 512 gigabytes of RAM to store intermediate data without input-output bottlenecks psc.edu.

Benchmark Accuracy and Simulation Scope

The surrogate model was tested across eight neurite geometries alongside four abnormal transport scenarios psc.edu. The system accelerated calculation speeds by a factor of 10 compared to conventional methods while maintaining an average prediction error of approximately 3 percent relative to full-scale physics simulations, with peak errors remaining below 8 percent psc.edu.

Results appeared in Computer Methods in Applied Mechanics and Engineering in December 2025 psc.edu. Current limitations of GALDS include spatial averaging that bypasses local microscale mechanics and a restriction to static geometries, preventing the simulation of dynamically growing arborizations psc.edu. Ongoing development focuses on variable-resolution models capable of accommodating morphological changes psc.edu.

CMU Engineers Model Nerve Cell Transport 10 Times Faster on Bridges-2

Computational Performance on Bridges-2

Benchmark Accuracy and Simulation Scope

Sources · 2

Read also

UC San Diego Tests Ketamine-Assisted Mindfulness for Opioid Use Disorder in Phase 2 Trial