Abstract

Physics-based simulations analyze interactions between physical objects, especially in areas such as terramechanics to study the interaction between soil particles. By modern standards, most computer simulations run at 60 Hz or 120 Hz refresh rates, equating to real-time step size require- ments of 16 ms or 8 ms, respectively. The discrete element method (DEM) models the interactions between the soil particles accurately, but it requires a lot of computational power. Graph Network Simulator (GNS) presents a promising alternative where it uses graph neural networks to learn and approximate the dynamics of the physical system fed into it. Although this method is faster than conventional models, the high dimensionality of the dataset means that the inference happens on a very large graph and rollouts do not occur in real-time. To accelerate the simulation, Haeri and Skonieczny (2021) developed a method that uses dimensionality reduction techniques such as principal component analysis (PCA). This method identifies the top principal components or PCA modes in the dataset that carry the highest variance. The next step is to feed these modes into the GNS along with the rigid body to train the model. Even though the reduced-order dataset results in a much smaller graph than the original dataset, the modified real-time GNS does not use algorithms such as nearest neighbours during adjacency graph construction for training and inference because the interaction between the rigid body and PCA modes may not be proximity-based. The graph fed into the network is naively fully connected. The contribution of this research is to propose a partial graph framework for the GNS to further accelerate this subspace framework without compromising the performance. To identify the redundant connections in the adjacency graph, Neural Relational Inference (NRI) is used. The NRI model, based on a variational autoencoder, uses the encoder to extract a partial graph between the PCA modes. This graph contains the most important edges that can still be used to perform inference on the GNS without any significant loss in accuracy. The effectiveness of this approach has been tested on blade cutting as well as wheel datasets to generate simulation at 60 Hz. This framework has reduced the inference time for excavation blade-driven granular flow from (0.34 - 0.48) seconds per second of simulation to (0.12 - 0.15) seconds and from 0.37 seconds to 0.18 seconds for wheel driven granular flow, achieving approximately 3x and 2x speed up respectively.