Optimal parallelization strategies for active flow control in deep reinforcement learning-based computational fluid dynamics DOI
Wang Jia, Hang Xu

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(4)

Published: April 1, 2024

Deep reinforcement learning (DRL) has emerged as a promising approach for handling highly dynamic and nonlinear active flow control (AFC) problems. However, the computational cost associated with training DRL models presents significant performance bottleneck. To address this challenge enable efficient scaling on high-performance computing architectures, study focuses optimizing DRL-based algorithms in parallel settings. We validate an existing state-of-the-art framework used AFC problems discuss its efficiency bottlenecks. Subsequently, by deconstructing overall conducting extensive scalability benchmarks individual components, we investigate various hybrid parallelization configurations propose strategies. Moreover, refine input/output (I/O) operations multi-environment to tackle critical overhead data movement. Finally, demonstrate optimized typical problem where near-linear can be obtained framework. achieve boost from around 49% approximately 78%, process is accelerated 47 times using 60 central processing unit (CPU) cores. These findings are expected provide valuable insight further advancements studies.

Language: Английский

Principles of artificial intelligence in radiooncology DOI Creative Commons
Yixing Huang, Ahmed M. Gomaa,

Daniel Höfler

et al.

Strahlentherapie und Onkologie, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 6, 2024

Abstract Purpose In the rapidly expanding field of artificial intelligence (AI) there is a wealth literature detailing myriad applications AI, particularly in realm deep learning. However, review that elucidates technical principles learning as relevant to radiation oncology an easily understandable manner still notably lacking. This paper aims fill this gap by providing comprehensive guide specifically tailored toward oncology. Methods light extensive variety AI methodologies, selectively concentrates on specific domain It emphasizes principal categories models and delineates methodologies for training these effectively. Results initially distinctions between well supervised unsupervised Subsequently, it fundamental major models, encompassing multilayer perceptrons (MLPs), convolutional neural networks (CNNs), recurrent (RNNs), transformers, generative adversarial (GANs), diffusion-based reinforcement For each category, presents representative alongside their Moreover, outlines critical factors essential such data preprocessing, loss functions, optimizers, other pivotal parameters including rate batch size. Conclusion provides overview enhance understanding AI-based research software applications, thereby bridging complex technological concepts clinical practice

Language: Английский

Citations

5
Recent advances in culture medium design for enhanced production of monoclonal antibodies in CHO cells: A comparative study of machine learning and systems biology approaches DOI

Hossein Kavoni,

Iman Shahidi Pour Savizi, Nathan E. Lewis

et al.

Biotechnology Advances, Journal Year: 2024, Volume and Issue: unknown, P. 108480 - 108480

Published: Nov. 1, 2024

Language: Английский

Citations

5
Dynamic warning zone and a short-distance goal for autonomous robot navigation using deep reinforcement learning DOI Creative Commons
Estrella Montero, Husna Mutahira, Nabih Pico

et al.

Complex & Intelligent Systems, Journal Year: 2023, Volume and Issue: 10(1), P. 1149 - 1166

Published: Aug. 22, 2023

Abstract Robot navigation in crowded environments has recently benefited from advances deep reinforcement learning (DRL) approaches. However, it still presents a challenge to designing socially compliant robot behavior. Avoiding collisions and the difficulty of predicting human behavior are crucial challenging tasks while navigates congested social environment. To address this issue, study proposes dynamic warning zone that creates circular sector around humans based on step length speed humans. properly comprehend keep safe distance between humans, zones implemented during robot’s training using enforcement techniques. In addition, short-distance goal is established help efficiently reach through reward function penalizes for going away rewards advancing towards it. The proposed model tested three state-of-the-art methods: collision avoidance with (CADRL) , long short-term memory (LSTM-RL), attention (SARL). suggested method Gazebo simulator real world operating system (ROS) scenarios. first scenario involves attempting free space. second uses static obstacles, third experimental results demonstrate performs better than previous methods leads an efficient time.

Language: Английский

Citations

13
AI-driven photonics: Unleashing the power of AI to disrupt the future of photonics DOI Creative Commons
Mohamed Mahmoud,

Amr S. Hares,

Mohamed Farhat O. Hameed

et al.

APL Photonics, Journal Year: 2024, Volume and Issue: 9(8)

Published: Aug. 1, 2024

Recent advances in artificial intelligence (AI) and computing technologies are currently disrupting the modeling design paradigms photonics. In this work, we present our perspective on utilization of current AI models for photonic device design. Initially, through physics-informed neural networks (PINNs) framework, embark task modal analysis, offering a unique networks-based solver utilizing it to predict propagating modes their corresponding effective indices slab waveguides. We compare model’s predictions against theoretical benchmarks finite differences solver. Evidently, using 349 analysis points, PINN approach had relative percentage error 0.69272% compared method, which 1.28142% with respect analytical solution, indicating that was more accurate conducting analysis. Our continuity over entire solution domain enhances its performance flexibility while requiring no training data due guidance by Maxwell’s equations, setting apart from most approaches. model also flexibly enables simultaneous prediction multiple any specified intervals indices. addition, novel reinforcement learning (RL)-based paradigm, employing an actor–critic inverse utilize paradigm optimize transmittance grating coupler manipulating geometry. Comparing obtained Particle Swarm Optimization (PSO) algorithm, RL-based effectively produced significant enhancement 34% 14 iterations only initial PSO, prematurely scored 27% 30 iterations, proving navigates space efficiently, achieving better than PSO resulting superior Based these approaches, discuss future photonics forward untapped potential bringing worlds together.

Language: Английский

Citations

5
Optimal parallelization strategies for active flow control in deep reinforcement learning-based computational fluid dynamics DOI
Wang Jia, Hang Xu

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(4)

Published: April 1, 2024

Deep reinforcement learning (DRL) has emerged as a promising approach for handling highly dynamic and nonlinear active flow control (AFC) problems. However, the computational cost associated with training DRL models presents significant performance bottleneck. To address this challenge enable efficient scaling on high-performance computing architectures, study focuses optimizing DRL-based algorithms in parallel settings. We validate an existing state-of-the-art framework used AFC problems discuss its efficiency bottlenecks. Subsequently, by deconstructing overall conducting extensive scalability benchmarks individual components, we investigate various hybrid parallelization configurations propose strategies. Moreover, refine input/output (I/O) operations multi-environment to tackle critical overhead data movement. Finally, demonstrate optimized typical problem where near-linear can be obtained framework. achieve boost from around 49% approximately 78%, process is accelerated 47 times using 60 central processing unit (CPU) cores. These findings are expected provide valuable insight further advancements studies.

Language: Английский

Citations

4