Learning continuous-time PDEs from sparse data with graph neural networks

Authors: Valerii Iakovlev, Markus Heinonen, Harri Lähdesmäki

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our model s performance in learning the dynamics of known physical systems. We compare to state-of-the-art competing methods, and begin by performing ablation studies to measure how our model s performance depends on measurement grid sizes, interval between observations, irregular sampling, amount of data and amount of noise.
Researcher Affiliation Academia Valerii Iakovlev, Markus Heinonen & Harri Lähdesmäki Department of Computer Science Aalto University Helsinki, Finland {valerii.iakovlev, markus.o.heinonen, harri.lahdesmaki}@aalto.fi
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Scripts and data for reproducing the experiments can be found in this github repository.
Open Datasets No The paper states that training data was obtained by solving initial-boundary value problems and downsampling these solutions. It does not provide concrete access information (e.g., a link, DOI, or formal citation with authors/year) for a publicly available dataset.
Dataset Splits No The paper specifies the training and testing data sizes but does not explicitly mention a validation data split or its details.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions “torchdiffeq Python package” and “Rprop optimizer” but does not provide specific version numbers for these software components or the Python language.
Experiment Setup Yes The model used for all following experiments contains a single graph layer. The mean was selected as the aggregation function. Functions φ(1)(ui, ) and γ(1)(ui, uj ui, xj xi) were represented by multilayer perceptrons with 3 hidden layers and hyperbolic tangent activation functions. Input/output sizes for φ(1) and γ(1) were set to 4/40 and 41/1 respectively. The number of hidden neurons was set to 60. This gives approximately 20k trainable parameters. [...] adaptive-order implicit Adams solver was used with rtol and atol set to 1.0 10 7. Rprop (Riedmiller & Braun, 1992) optimizer was used with learning rate set to 1.0 10 6 and batch size set to 24.