Models Overview

RealPDEBench benchmarks 10 baseline models spanning classical reduced-order methods, neural operators, transformers, and foundation models.

This page summarizes what is included and the shared model interface. For detailed architecture notes and RealPDEBench-specific constraints, see the per-model pages.

Common interface in RealPDEBench

All learning baselines implement a unified Model API:

  • forward(x) -> y: predicts future spatiotemporal fields
  • train_loss(input, target): returns a training loss used by the benchmark training loop

Tensor layout (shared across baselines):

  • A dataset sample returns (input, target) with shapes [T_in, H, W, C_in] and [T_out, H, W, C_out].
  • After batching, the model sees \(x \in \mathbb{R}^{B \times T_\text{in} \times H \times W \times C_\text{in}}\) and must output \(y \in \mathbb{R}^{B \times T_\text{out} \times H \times W \times C_\text{out}}\).

Practical notes:

  • Many baselines vectorize multi-step prediction and assume \(T_\text{out}\) is a multiple of \(T_\text{in}\) (often \(T_\text{out}=T_\text{in}\)).
  • For partially observed real-world modalities, unmeasured channels are represented by zero padding; simulated training uses mask-training to reduce train–test mismatch.

Baseline roster

Family Baseline Key reference
Traditional DMD Kutz et al., Dynamic Mode Decomposition, 2016
Trad & CNN U-Net Ronneberger et al., MICCAI 2015
Neural Operators CNO Raonic et al., NeurIPS 2023
Neural Operators DeepONet Lu et al., Nature Machine Intelligence, 2021
Neural Operators FNO Li et al., ICLR 2021
Neural Operators MWT Gupta et al., NeurIPS 2021
Transformers GK-Transformer Cao, NeurIPS 2021
Transformers Transolver Wu et al., ICML 2024
Neural Operators WDNO Hu et al., ICLR 2025
Foundation Models DPOT (DPOT-S / DPOT-L) Hao et al., ICML 2024

Category pages

Add your own model

To implement and register a new model in the benchmark codebase, see: Add Your Model →

Where to find results

Benchmark numbers are shown in the homepage Results Explorer: Jump to Results Explorer.