Differentiable Fluids with Solid Coupling for Learning and Control

Authors: Tetsuya Takahashi, Junbang Liang, Yi-Ling Qiao, Ming C. Lin6138-6146

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

Reproducibility Variable Result LLM Response
Research Type Experimental We implemented our differentiable fluid simulator in C++ and integrated it with deep neural networks implemented in Py Torch 1.5 (Paszke et al. 2019). To evaluate the effectiveness of our method, we conduct three types of experiments on an Intel Core i5-7200U with 8GB RAM. First, we perform an ablation study to evaluate the performance gain compared to other gradient computation techniques in 5.1. Then, we apply our method to optimization problems in 5.2 and control problems 5.3. More details and results can be found in the supplementary material.
Researcher Affiliation Collaboration Tetsuya Takahashi1, 2, Junbang Liang2, Yi-Ling Qiao2 and Ming C. Lin2 1Adobe, 2University of Maryland at College Park
Pseudocode Yes Algorithm 1 Gradient computation with the adjoint method 1: Store q0 2: for i = 0, . . . , n 1 do 3: Forward simulation: qi+1 = Fi(qi, fi) 4: Store qi and intermediate results 5: rn = φ qn 6: for i = n 1, . . . , 0 do 7: Backward update: ri = Fi qi T ri+1 + φ qi 8: Compute gradient: dφ dfi = r T i Fi
Open Source Code No The paper mentions 'More details and results can be found in the supplementary material' but does not provide a direct link to the source code or explicitly state its release for the described methodology.
Open Datasets No The paper describes experimental scenarios and setups ('scene shown in Figure 2', 'solid control scenario, as shown in Figure 5'), but does not explicitly mention or provide access information (link, DOI, or formal citation with author/year) for any publicly available or open dataset used.
Dataset Splits No The paper refers to 'training' neural networks, but it does not specify any dataset splits (e.g., percentages, sample counts) for training, validation, or testing, nor does it refer to predefined splits from cited sources.
Hardware Specification Yes To evaluate the effectiveness of our method, we conduct three types of experiments on an Intel Core i5-7200U with 8GB RAM.
Software Dependencies Yes We implemented our differentiable fluid simulator in C++ and integrated it with deep neural networks implemented in Py Torch 1.5 (Paszke et al. 2019).
Experiment Setup Yes We use the simulation state and time as an input for the neural network, 30 and 20 nodes in the first and second layers, respectively, with the Re LU activation, and finally generate three dimensional control forces for controlling the orange box.