POSEIDON
New1. Overview
POSEIDON is a PDE foundation model and downstream fine-tuning suite released in 2024 by the Computational and Applied Mathematics (CAM) laboratory at ETH Zurich. Its central claim is not another neural operator trained for a single equation, but a workflow in which a model pretrained on a broad set of PDE data is adapted to a new equation family with very few samples. Instead of training an FNO or DeepONet from scratch for every problem, you take a model pretrained on large and diverse PDE corpora and fine-tune it lightly to your own operator. For a physics-AI practitioner this changes the economics of operator learning: the cost shifts from a full training run per equation to one pretraining plus many cheap adaptations. This site compares POSEIDON with NeuralOperator and PhysicsNeMo because the three define "operator" and charge for it in completely different ways.
After reading, you should be able to judge whether, for a project that must repeatedly query a family of structurally similar but parametrically varied PDEs, pretraining plus fine-tuning is cheaper than training from scratch. The companion neural-operator selection guide and the neural-operators collection keep the comparison anchored to the deliverable. All capability statements below were verified against the project repository and the POSEIDON paper on 11 November 2026, and claims that are the project's own description rather than an independent measurement are labelled as such in Section 9.
2. How It Works
POSEIDON builds on an operator-learning backbone (a spectral or convolutional neural-operator architecture) and a data-efficiency trick the authors call All2All semigroup augmentation. The key observation is temporal separability: a single trajectory of length T contains far more training pairs than its raw length suggests, because any pair of time slices can serve as an input-output pair for the time-evolution operator. By exploiting this, the augmentation extracts on the order of O(T^2) training pairs from one trajectory, which is what lets the model learn from very few full simulations. A pretrained POSEIDON model is then queried like an operator: you supply a new initial condition or coefficient field and the model returns the solution field, including at resolutions not seen during training, which the project calls mesh-invariant inference.
The fine-tuning path is what makes it a foundation model rather than a fixed surrogate. Given a new PDE family, you load the pretrained weights and fine-tune on a small set of solutions (the project reports strong results with roughly twenty training samples for several tasks, against roughly a thousand for a comparable FNO trained from scratch). This keeps the data-generation cost low for the downstream user. The architecture and the augmentation are described in the POSEIDON paper (arXiv 2405.19101); the repository ships the pretrained checkpoints and the downstream benchmark spanning fifteen task families covering elliptic, parabolic, hyperbolic and mixed regimes, which is the practical evidence base for the data-efficiency claim.
3. Scope and Assumptions
POSEIDON assumes your problem is an operator: a mapping from an input field (initial condition, coefficient, geometry description) to an output field (the evolved or solved state), and that you can obtain at least a handful of full simulations for the target family to fine-tune on. It is designed for regular gridded or mesh-compatible fields where the operator structure is shared across queries. It is not a general-purpose solver that returns certified accuracy on arbitrary geometries, and it does not replace a classical solver on a single large high-fidelity query where raw throughput matters. Like every data-driven operator, its reliability is bounded by the distribution covered by pretraining and fine-tuning data.
The practical assumption is that you accept an amortised-accuracy model: the model is fast across many queries but its error must be measured on held-out configurations rather than assumed. The licence is whatever the repository LICENSE states, and the pretrained weights carry their own usage terms, so any commercial deployment should read both before shipping. A mismatch between the problem you hand it and the operator contract it satisfies is the most common source of disappointment, and the section on failure modes returns to this with concrete examples. POSEIDON is a tool for repeated, similar queries, not a one-off certified solve.
4. Getting Started
The fastest entry is the repository's downstream benchmark and the provided pretrained checkpoints: clone the CAM-lab POSEIDON repository, install the dependencies, and run the documented fine-tuning scripts on one of the fifteen benchmark families to see a working configuration before touching your own data. A first experiment follows the documented workflow: load a pretrained checkpoint, prepare a small set of input-output pairs for your equation family, run the fine-tuning script, and evaluate on a held-out configuration with an independent reference. The official examples and the paper walk through the elliptic and advection-dominated cases as the gentlest path rather than starting from an empty model. The most frequent beginner failure is fine-tuning on too few and too similar samples, which makes the model look good in-sample and fail out-of-distribution.
Before trusting a result, establish a held-out reference on a configuration the fine-tuning data did not cover, because the model's value is exactly its behaviour away from the training pairs. Record the pretrained checkpoint version, the fine-tuning sample count, the resolution and the evaluation metric next to every result, because those four values decide whether a result is reproducible, and a result that cannot be reproduced cannot be defended to a customer. The companion physics-based deep-learning material and the operator-learning survey on this site supply the surrounding theory, and the PINN landscape guide organises when an operator model is worth the compute.
5. Comparison with Alternatives
The useful comparison is by contract. NeuralOperator provides trained operator architectures such as FNO that you assemble and train yourself, while POSEIDON ships a pretrained foundation model that you adapt; the training cost therefore sits in different places. PhysicsNeMo is a large training framework with many physics-specific models and pipelines, tuned for scale and enterprise integration, whereas POSEIDON is a focused foundation model plus a downstream benchmark. DeepXDE trains a network from scratch for the equation you provide via a residual formulation, which is a different economic model from pretrain-and-adapt. JAX-Fluids and OpenFOAM are differentiable or classical solvers rather than operator surrogates.
Within operator learning, POSEIDON's differentiator is the All2All semigroup augmentation and the pretrain-and-adapt economics: it targets the regime where you must query a family of similar equations many times. A classical reduced-order model or an FNO trained from scratch can also amortise repeated queries, but each requires its own training run and its own verified reference. The neural-operator selection guide keeps these comparisons honest by stating the deliverable: a certified single solve, an interactive design sweep, or a surrogate that amortises thousands of queries. A framework that wins on data efficiency may lose where certified accuracy or a single high-fidelity solve dominates.
6. Benchmarks and Evidence
Public evidence is the project's own downstream benchmark and the POSEIDON paper. The paper reports that the All2All semigroup augmentation yields order-of-magnitude more training pairs per trajectory, and that fine-tuned POSEIDON reaches competitive accuracy on several of the fifteen benchmark families with roughly twenty training samples, whereas an FNO trained from scratch on the same families needs on the order of a thousand. These are project-authored or community capability statements tied to the published benchmark, not independent blind validations of accuracy or speed at scale, and they are measured on the benchmark's own train-test splits. The benchmark covers a deliberately broad set of regimes so the claim is about breadth of transfer, not a single best number.
Independent context tempers the optimism. Operator-learning models in general inherit the limits of their training distribution, and work on benchmark suites such as PDEBench and The Well shows that cross-distribution generalisation is the hard part. A 2024 review finding that many "ML beats classical solvers" papers compared against weak baselines is a reminder that any accuracy figure from POSEIDON should be re-measured on an independent reference for your specific family. These are not reasons to reject POSEIDON; they are reasons to treat any result as unverified until the checkpoint, sample count, resolution and hardware are stated, following the same discipline used by the aerospace-CFD case and the automotive-aero case on this site.
7. Role in Physics-AI Workflows
In a production pipeline POSEIDON sits at the amortised-query end. It is most valuable when you must evaluate a family of similar PDEs thousands of times - design sweeps, parametric uncertainty propagation, or online control where a fast surrogate stands in for a classical solver inside a validated envelope. Its credibility rests entirely on how thoroughly the fine-tuned model was validated against held-out classical or experimental references; a pretrained operator is only as good as the last independent solve you ran on a configuration it had not seen. Benchmarks such as PDEBench and The Well are examples where classical and neural solvers provide the ground truth that surrogates learn from.
POSEIDON also fits into a longer verification chain. A classical solver remains the reference for correctness and for certification claims; a physics-informed model such as DeepXDE can supply residual corrections where labels are scarce; and a foundation-style model such as POSEIDON can provide a pretrained starting point for operator learning. None of these removes the need for an independent reference solve on held-out configurations. The physics-AI learning path on this site places these roles in sequence, and the surrogate-modeling-rom collection collects the underlying entry points. Writing down the expected query count before fine-tuning is still the cheapest way to avoid an unrecoverable data-generation investment.
8. Limitations and Failure Modes
POSEIDON inherits every limitation of data-driven operator learning. Its accuracy is bounded by the distribution covered by pretraining and fine-tuning data; a query far from that distribution yields smooth, plausible and wrong fields, and the model will not flag that for you. Fine-tuning on too few or too similar samples produces in-sample success and out-of-distribution failure, which is the most common practical pitfall. Sharp discontinuities, shocks and multi-scale features are not automatically resolved just because the model is a foundation model; the architecture still smears what the data did not teach it to represent.
There are practical limits too. POSEIDON is a research artifact centred on regular grids and the published benchmark families, so adapting it to irregular industrial meshes or to equations absent from the pretraining mix needs care and extra validation. The licence is whatever the repository LICENSE states and the pretrained weights carry their own terms, so read both before any commercial use; this site reports the licence as "see repository" rather than asserting a specific one. POSEIDON does not provide certified accuracy or an audit trail on its own; when the deliverable must be defended, a classical solve or an experiment stays in the loop as the reference. Treating a surrogate's smooth output as ground truth far from the training distribution is a modelling error, not a tool failure.
9. Version and Source Notes
Record 1: POSEIDON was published by the ETH Zurich CAM laboratory, with the paper "Poseidon: Efficient Foundation Models for PDEs" on arXiv as 2405.19101 (29 May 2024); the repository is camlab-ethz/POSEIDON. Record 2: the project reports that All2All semigroup augmentation extracts on the order of O(T^2) training pairs per trajectory, and that fine-tuned POSEIDON reaches competitive accuracy on several benchmark families with roughly twenty training samples versus roughly a thousand for an FNO trained from scratch; these are project-authored benchmark figures. Record 3: the downstream benchmark spans fifteen task families across elliptic, parabolic, hyperbolic and mixed regimes, and the model supports mesh-invariant inference at resolutions beyond those seen in training.
Record 4 (reconciliation discipline): when citing a POSEIDON result, the pretrained checkpoint version, the fine-tuning sample count, the resolution and the evaluation metric must be recorded together, because the model and its error depend on exactly those choices. An accuracy figure quoted without its checkpoint and sample count cannot be reproduced and should not be carried into a proposal. The licence is "see repository LICENSE" and the pretrained weights carry their own terms; any deployment should read both. All records were verified on 11 November 2026 against the project repository and the POSEIDON paper; vendor-stated benchmark figures are treated as project-authored capability statements rather than independent validations.
10. References
- Herde M., Raonic B., Rohner T., Käppeli R., Molinaro R., de Bézenac E., Mishra S., "Poseidon: Efficient Foundation Models for PDEs", NeurIPS 37 (2024), arXiv:2405.19101.
- The POSEIDON Project, "POSEIDON repository", GitHub, github.com/camlab-ethz/POSEIDON (accessed 11 November 2026).
- Li Z. et al., "Fourier Neural Operator for Parametric Partial Differential Equations", ICLR 2021, arXiv:2010.08895.
- Lu L. et al., "DeepONet: Learning nonlinear operators for identifying differential equations", Nature Machine Intelligence 3, 218-229 (2021).
- Takamoto M. et al., "PDEBench: An Extensive Benchmark for Scientific Machine Learning", NeurIPS 2022, arXiv:2210.07182.
- Ohana R. et al., "The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning", NeurIPS 2024 Datasets and Benchmarks, arXiv:2412.00568.
- Kovachki N. et al., "Neural Operator: Learning Maps Between Function Spaces", Journal of Machine Learning Research 24, 1-97 (2023), arXiv:2108.08481.
- Gupta G., Brandstetter J., "Towards Multi-spatiotemporal-scale Generalized PDE Modeling", arXiv:2209.15616 (2022).
- Bartolucci F., de Bézenac E., Raonic B., Molinaro R., Mishra S., Alaifari R., "Representation Equivalent Neural Operators: a Framework for Alias-free Operator Learning", NeurIPS 36 (2023), arXiv:2305.19913.
- Raonic B. et al., "Convolutional Neural Operators for robust and accurate learning of PDEs", NeurIPS 36 (2023), arXiv:2302.01178.