Framework for distributed computation of GPU-based MC simulations to build a large machine learning skin training data set (Conference Presentation)

We showcase a large computation job distribution framework to expedite the execution of GPU-intensive simulations called The Offloader. This approach leverages parallel computing by breaking down a large set of simulations into manageable sub-tasks, which are then distributed across multiple computers in the Microsoft Azure cloud, giving a linear reduction in simulation times as the number of available computers increases. We integrate this capability with a machine learning (ML) algorithm that dynamically selects informative points within the free parameter space of a skin tissue model. With each new simulation, we enrich the training data, increasing the ML model’s accuracy. The Offloader's speed with our Machine Learning framework enables optimization across various parameters, including source-detector distance, incident light wavelength, and detector geometries allowing us to explore these dimensions efficiently for the development of new or improved optical devices and measurement techniques.