FunDiff
收藏资源简介:
FunDiff是一个新型框架,用于在函数空间中进行生成建模,旨在解决物理学中生成连续函数数据的挑战。该框架结合了潜在扩散过程和函数自动编码器架构,能够处理具有不同离散化的输入函数,生成可在任意位置评估的连续函数,并自然地融入物理先验。这些先验通过架构约束或物理信息损失函数来实施,确保生成的样本满足基本物理定律。FunDiff在流体动力学和固体力学领域的应用中显示出其实用效果,能够生成与目标分布高度一致且对噪声和低分辨率数据具有鲁棒性的物理一致样本。代码和数据集可在提供的网址中公开获取。
FunDiff is a novel framework for generative modeling in function spaces, designed to address the challenge of generating continuous functional data in physics. This framework combines latent diffusion processes and functional autoencoder architectures, enabling it to handle input functions with varying discretizations, generate continuous functions that can be evaluated at arbitrary locations, and naturally integrate physical priors. These priors are enforced through architectural constraints or physics-informed loss functions, ensuring that the generated samples adhere to fundamental physical laws. FunDiff has demonstrated its practical efficacy in applications across fluid dynamics and solid mechanics, capable of generating physically consistent samples that align closely with the target distribution and are robust to noise and low-resolution data. Code and the dataset are publicly available at the provided URL.




