(216c) Generic Universal Modeling Platform Software

We present the Generic Universal Modeling Platform Software (GUMPS) as a modeling support library. The purpose of this software is to aid modelers with common tasks while also taking care to follow best practices in their implementation. This is accomplished by dividing a model into smaller component pieces: a Kernel, a Study, a Solver and a Problem. Surrogate model building only requires a working model with a vector of inputs and outputs and the library takes care of the sampling, variable transformations, model training, and prediction. Uncertainty Quantification can use a surrogate or a model and only requires the target probabilities and degree of accuracy required and will sample until the requirements are met. The ODE solver can handle discontinuities with the modelers only needing to specify when they have a discontinuity in the system. Parameter estimation and optimization are supported with both single and multiple-objectives with a composable system of loss functions and automatic variable transformation in order to improve convergence with visualization and aid in understanding.

The GUMPS software is written in Python and has full unit test coverage. We use parallel processing on CPU and GPU as appropriate to the problem. Many of the advanced features are built using pyTorch.