Join MindSpore Special Interest Groups for AI Fluid Simulation Development

Computational Fluid Dynamics (CFD) is fundamental to the research and development of aerospace, marine equipment, energy, and power.

However, it faces several challenges and bottlenecks. For instance, the mesh generation for fluid dynamics is complex and cannot be fully automated at complex boundaries. Simulations rely on complex iterative computations, whose speedup is limited during parallel computing. It is also challenging to solve high-dimensional equations while maintaining both precision and performance.

These challenges present new opportunities for AI to penetrate scientific computing. AI methods do not rely on mesh generation. Instead, they have inherent parallel inference capabilities and can deliver computing results without going through iterative computing. AI methods are quick in learning the laws of physics and presenting results through inference, while striking a satisfactory balance between precision and performance. Major companies such as Google and NVIDIA have already started applying AI to fluid dynamics research.

Unlike the computer vision (CV) and natural language processing (NLP) fields, where there are stable and user-friendly model libraries and communities, the AI fluid dynamics field lacks such resources. This makes it difficult for algorithm researchers to experiment with algorithms and for application developers to train and deploy models.

Therefore, the establishment of the MindFlow Special Interest Group (MindFlow SIG) and the invitation for like-minded partners from the open source community to join, is aimed at promoting the development of fluid dynamics research, as well as facilitating the application and innovation of AI. This can be achieved by creating an open, user-friendly, and efficient AI fluid dynamics library and communication community.

1. Introduction to MindFlow SIG

The SIG is dedicated to MindSpore Flow, leveraging its advantages to enhance the functionality of the toolkit, grow the community ecosystem, and deliver efficient and user-friendly AI computational fluid simulation tools to researchers, educators, and students. It also serves as a platform for collaboration and communication among influential and interested individuals.

2. Our Mission

Our focus is on applying AI in fluid dynamics and exploring multiple AI fluid dynamics simulation paradigms, including physics-based simulation, data-driven simulation, and data and mechanism integrated simulation. The goal is to build a simple and efficient AI+fluid computing framework that supports research in AI flow fields for aerospace, marine hydrodynamics, energy, and power.

The MindFlow SIG has the following key tasks:

2.1 Physics-based AI Fluid Simulation

Physics-based AI fluid simulation involves incorporating a physics equation into a neural network's loss functions to train the equation and ensure that the learning outcome adheres to the laws of physics. This module is mainly used for solving partial differential equations (PDEs), data fusion inverse problems, and data assimilation. MindSpore Flow currently offers common boundary and initial condition setting methods, sampling methods, and loss function construction to solve and optimize classical fluid dynamics problems.

2.2 Data-driven AI Fluid Simulation

Data-driven AI fluid simulation relies on a large amount of fluid simulation data. A delicately designed neural network can mine the laws of physics between data samples, enabling efficient, parallel, and fast inference. In addition, neural operator learning, such as Fourier neural operator (FNO) and deep operator network (DeepONet), has a certain parameter generalization capability. This module is suitable for scenarios that require quick inference and parameter space design optimization with a large amount of labeled data. Currently, MindSpore Flow offers network models like Vision Transformer (ViT) and FNO, which can solve classical fluid flow problems (e.g., Burgers' equation) and engineering flow problems.

2.3 Data and Mechanism Integrated AI Fluid Simulation

The data and mechanism integrated AI fluid simulation, represented by PDE-Net, can accurately predict the dynamic characteristics of complex systems by learning partial differential equations from data. PDE-Net is particularly used in scenarios where there are limited scientific data samples and the control equation is known. The built-in flow field equation information reduces the data volume required by the neural network and improves the network generalization. Additionally, coupling differentiable Computational Fluid Dynamics (CFD) solver with neural networks enables AI correction, AI frame interpolation, and AI super-resolution. Currently, MindSpore Flow offers the basic capability of PDE-Net, which can solve classical fluid flow problems.

2.4 Differentiable CFD Solver

Fluid simulation software mainly solves the control equation of fluid dynamics in a computer using numerical methods. This allows for the analysis, prediction, and control of fluids, making it a valuable tool in industries such as aerospace, ship manufacturing, energy, and power. The differentiable CFD solver, based on MindSpore Flow, offers several advantages, including Just-in-time (jit) compilation, Vectorised Object Mapping (vmap) automatic vectorization, Autograd end-to-end automatic differentiation, and support for various hardwares.

3. Framework and Code Repository

MindSpore Flow is a collection of fluid simulation tools developed on the MindSpore platform. It enables AI-based flow field simulations in aerospace, marine hydrodynamics, energy, and power. Its goal is to offer efficient and user-friendly AI computing fluid simulation software to researchers, engineers, teachers, and students in the industry. MindSpore Flow offers a range of common functions, including physics-based simulation, data-driven simulation, and data and mechanism integrated simulation.

4. Our Plan

4.1 Early Stage

We will host academic communication for members, providing references for the evolution and improvement of MindSpore Flow's functions. Specifically, we will organize one large and several small activities each year, as well as a campus activity every quarter. Additionally, the SIG will hold a major summer school activity annually, inviting core experts and teachers to teach on various topics for multiple days. The team teacher will guide members in conducting scientific and technological research, expanding code repository functions, and fixing bugs. Members can also utilize the MindSpore Flow software for their own research and development. The group will release open source internships and crowd intelligence tasks in the community for students and teachers to participate in.

4.2 Later Stage

We will carry out collaborative research within the community through cooperative development and other modes to facilitate the implementation of a wider range of applications.

5. Organizational Structure

Team members:

Maintainer: hsliu_ustc, MindSpore Senior Engineer

Maintainer: zwdeng, MindSpore Senior Engineer

Maintainer: Yi_zhang95, MindSpore Senior Engineer

Member: hong-ye-zhou, MindSpore Senior Engineer

Member: Bokai Li, MindSpore Engineer

Member: liulei277, MindSpore Engineer

Member: yangge_nihilism, MindSpore Engineer

Member: haojiwei, postgraduate student at Beihang University

Member: Li Zhengyi, doctoral student at Peking University

Member: Li Zhuoyuan, doctoral student at Peking University

Member: Ye Zhanhong, doctoral student at Peking University

Member: Wang Qi, doctoral student at Renmin University of China

Member: Lei Yuxiang, undergraduate at Wuhan University

MindFlow SIG Calls for You.

SIG repository:

https://gitee.com/MindSpore/community/tree/master/sigs/mindflow

Code repository:

https://gitee.com/mindspore/mindscience/tree/master/MindFlow

The MindSpore Community warmly invites industry experts and academic partners to establish special interest groups (SIGs) within the community. Let's work together to cultivate a technical identity within the field and foster an open-source MindSpore ecosystem.

The MindSpore community has established dozens of SIGs to create an open communication platform for experts, professors, students, and enthusiasts in the field. The purpose is to promote technical exchange and academic communication through activities such as meetings and project development, and to enhance the influence and technological capabilities of SIG members.