Performing Department
(N/A)
Non Technical Summary
As the demand for fish increases and the aquaculture systems expand to deeper oceans, traditional manual operations and monitoring are becoming less effective. This drives the urgent need to develop and deploy intelligent autonomous systems like digital twin models, which are virtual representations of real aquaculture systems. The success of the project can help make aquaculture more automated and more sustainable. To achieve this goal, we will build fast machine-learning models that can accurately present the dynamics of aquaculture systems in different environmental and operation states. The data to train the models comes from the hydrodynamics simulations of aquaculture systems. The comparison of the predictions from the machine learning model and real-time sensor measurements can guide the field operators to take immediate control actions and prevent potential system failures.We aim to release the datasets and predictive models and present the work in journals, conferences, and workshops in aquaculture engineering and machine learning. We will also create a graduate-level course on sustainable aquaculture at Arizona State University. The success of the project will revolutionize the intelligent and precise operations of aquaculture systems and will prevent huge economic losses due to undesirable weather and faulty operations.
Animal Health Component
20%
Research Effort Categories
Basic
50%
Applied
20%
Developmental
30%
Goals / Objectives
This proposal's research objective is to design digital twin systems (reliable virtual models of complex systems) to predict the structural damage of offshore aquaculture systems under different environments and operating conditions. The success of the project will revolutionize the intelligent and precise operations of aquaculture systems, and will prevent huge economic losses due to undesirable weather and faulty operations. In this seed grant stage, we will conduct finite element simulations and computational fluid dynamics simulations to generate the aquaculture systems' response data in different environments. The collected data, in the form of videos, will be used to train graph neural network models. The machine learning models are expected to perform accurate dynamics predictions by learning on the computational meshes. The models will also accelerate the traditional mesh-based simulations in near-real-time, by passing nonlinear information between neighboring computational nodes. Due to the differentiable computations, the machine learning models can be used to estimate the damage and changes in the system properties quickly, given the measurement of aquaculture systems' dynamical response. Meanwhile, we will test the robustness of the model by introducing noise in the data and improve its robustness by incorporating physics-based laws. We will validate the machine learning models by conducting independent lab-scale tank experiments, and collaborating with industry.
Project Methods
The dynamics of aquaculture cages in ocean currents and waves are very high dimensional and nonlinear. This makes the traditional finite element or finite volume-based simulations extremely slow and expensive. To accelerate the simulation while learning the high-dimensional nonlinear dynamics of aquaculture systems, we will develop a physics-aware digital twin model based on a graph neural network technique.Meanwhile, asuccessful digital twin model also requires sensing technologies that provide feedback on the actual state. We will design a cost-constrianed machine learning model to determine the optimal sensor placement strategies.We will embed the fast digital twin models in a feedback control loop. Because of the digital twin system's differentiable nature, a cheap gradient-based optimization can be used. The aquaculture operators can then select the optimal control strategy to ensure system safety in different environmental conditions.