SPATIOTEMPORAL OPTIMIZATION
v20250624.1
One use of this framework is for optimization. Whether for a thermal energy network, power grid upgrades, carbon sequestration sites, transportation networks, or telecommunications networks, there are usually problems involving optimization of multiple variables at geographically distributed sites. These variables may vary over either space or time.
We demonstrated our framework on a 1 million node network on CPU, using the 4 variables of latency, throughput, processing delay, and reliability. Optimization meant minimizing latency and processing delay while maximizing throughput and reliability. The results: a relative upgrade priority for each node relative to the others so the entire network can reach near-peak performance in the shortest time.
Key differentiators
1) Including the overhead of file inputting/outputting and plotting the first 1000 nodes, this network was optimized in 2 minutes on CPU, with the custom measure necessary for the efficient representation used requiring 0.28 seconds to calculate. GPU speedups are typically 10-100x that of the CPU runtime due to efficient data parallelization.
2) Unlike optimization algorithms that will experience significant diminishing returns when a similarity measure nears a target value, only 2 cycles were required with our custom measure to reach 99.9% of the best value for each variable. In the context of neural networks for deep learning, this problem is known as a vanishing gradient.
WAVE CONVERSIONS AND MODELING
v20250525.1
Other uses of this framework are in the area of dynamics analysis, involving the more efficient modeling of waves, fields, networks, and flows. Expected applications are primarily in remote sensing (electromagnetic - satellite, aerial, UAS) and subsurface (seismic and EM) imaging, though research needs within signal processing and computational physics may also be served.
These slides provide examples of other framework capabilities to be extended.