Data Aggregator for Physics Informed Neural Networks in NVIDIA Modulus Framework

Resumo

The development of Physics Informed Neural Networks (PINNs) has been receiving considerable atten-
tion lately. PINNs incorporate the partial derivative equations that describe the physical behavior of a natural or

engineered system in the loss functions of neural networks. This model family represents a new paradigm for the

solutions of PDEs for both forward and inverse problems. Different frameworks that aim to facilitate the produc-
tion and training of such models are currently being provided, and Modulus is one of the available frameworks that

has been gaining ground recently. In any case, despite the capability of these packages to assist the construction of

PINNs, it is important to consider a viable data analysis strategy for the experiments. This work presents the Mod-
ulus Aggregator tool, which is developed to support the data analysis expert in the hyperparameter configuration

of multiple models produced, with a strategy for the aggregation of results. The aggregation tool complements the
TensorBoard visualization toolkit and takes advantage of the native directory structure of a Modulus experiment.
The experiment of a wave propagation shows the potential to assist the analysis of results and the possibility of
automating the extraction and filtering activities of trained models in a scenario of a significant amount of data.