Compact machine learning models for dynamical systems with Neural ODEs

"Ordinary differential equations lend themselves to describing the time evolution of natural and industrial systems. ODEs have been combined with modern machine learning techniques to create Neural ODEs. These gain the advantage of learning from big data in a familiar scientific framework. Their expressiveness allows for the encoding of prior knowledge while still being flexible enough to adapt to the training data. This allows us to capture nonlinear residual-physics that often elude easy characterization or are unknown, making the model more accurate and efficient to develop in the presence of data.

In this presentation , we discuss the advancements made in our implementation of the Neural ODE framework, its ability to learn and predict from tens of industrial strength benchmark datasets and pave way for developing models for some challenging datasets like those for Belgian pave tests for the automotive industry; something that continues to heavily rely on traditional field testing with limited simulation impact on durability and life prediction. In addition, we provide a framework suggestion on how these approaches and models can be deployed in various industries customer workflows.

"