Scientific Conditional Diffusion for Heat Fields¶
Notebook: ConditionalDiffusion_Heat2D_Tutorial.ipynb
This tutorial builds a conditional diffusion model for 2D heat fields. The task is a scientific field-reconstruction problem rather than a deterministic operator map: sparse sensor measurements of a smooth heat field are used to construct a conditional distribution over the full field.
Key ideas:
- the physical field comes from the 2D heat equation,
- conditioning is intentionally incomplete: sparse sensors plus a mask and a smooth inpainting baseline,
- predictive uncertainty is computed from the spread of generated samples,
- uncertainty should be largest away from sensors and larger on OOD fields than on in-domain fields.
The notebook shows:
- heat-field generation from an exact spectral heat solver,
- a reusable
ConditionalUNet2Ddenoiser, - DDPM-style training on residual fields,
- conditional sampling with data-consistency clamping at sensor points,
- sample mean, predictive standard deviation, and in-domain vs OOD comparisons.
Primary references:
- Ho, Jain, Abbeel (2020), Denoising Diffusion Probabilistic Models
- Song et al. (2021), Score-Based Generative Modeling through Stochastic Differential Equations
- Ronneberger, Fischer, Brox (2015), U-Net: Convolutional Networks for Biomedical Image Segmentation