Foundation ModelsMachine Learning PotentialsPretrainingFine-tuning

Physics Foundation Models

A collection of foundation models built for physical scenarios: universal machine-learning potentials, cross-domain pretrained foundation models with fine-tuning recipes, and multi-domain benchmarks plus the annual review.

Updated Nov 3, 2026 14 resources
Introduction

Why this collection

The foundation-model paradigm is entering physics — pretrain a base model, then fine-tune for a concrete task. That path is proven for text and images; in physics it is newer, but the direction is already clear.

This collection separates "physics foundation models" into three forms:

  • Universal machine-learning potentials (MLIPs) — replacing quantum-mechanics computation with a neural network. This is the most mature and most industrially valuable form: computing cost for materials, catalysis and batteries drops by orders of magnitude.
  • Cross-domain pretrained foundation models — pretrained on multi-physics data, then transferred to specific downstream tasks; Thermal-FM and GridSFM fall here.
  • Geometry and mesh foundation models — pretrained and generative models operating directly on geometry and meshes, with POSEIDON as a representative.

How this divides from the neural-operator collection: neural operators focus on learning solvers for a single PDE family, while foundation models focus on general representations across domains and tasks. In practice the two are merging.

For evaluation, look at The Well — one of the few multi-domain benchmarks built specifically for physical foundation models and directly citable for horizontal comparison.

Universal Machine Learning Potentials 5 resources

FAIR Chemistry and UMA: A Cross-Domain Atomic Foundation Model Platform

Materials and Chemistry / Atomistic Machine Learning
The consolidated repository of Meta FAIR Chemistry has grown from an OC20 electrocatalysis toolkit into a unified platform spanning molecules, materials, and catalysts, with the UMA model using a mixture-of-linear-experts architecture trained on 500 million structures and switching task names to serve seven domains from OC20 to OC25, OMat24, and OMol25.
New Repository code is MIT (Meta, Inc.); the fairchem-lammps package declares GnuV2 on PyPI; UMA weights are licensed separately under gated model terms; datasets carry their own terms Atomic Foundation ModelMachine Learning PotentialElectrocatalysis
💬 Editor note: The electrocatalysis screening case shows that universal potentials support material discovery pipelines, not just single-point energies.

MACE Machine-Learned Potentials: From Equivariant Message Passing to Atomistic Foundation Models

Molecular Dynamics / Machine-Learned Interatomic Potentials
An open-source equivariant graph-neural-network potential framework from the Csanyi group that unifies atomic cluster expansion with high-order equivariant message passing, reaching high accuracy with only two interaction layers, and now shipping several generations of foundation weights (MACE-MP, MH, OMOL, POLAR) for out-of-the-box molecular dynamics of crystals, liquids, gases, surfaces and reactive systems.
New MIT (code); some weights are released under the Academic Software License (ASL) for academic use only Molecular DynamicsMachine-Learned PotentialEquivariant Neural Network
💬 Editor note: MACE-MP-0 is a representative atomic-scale universal machine-learning potential, with published potentials covering many chemical systems.

FAIR Chemistry Machine Learning Toolkit

Materials / Chemistry / Framework Tutorial
Meta FAIR Chemistry open-source toolkit (formerly the Open Catalyst Project) spanning OC20/OC22 datasets and models like GemNet, eSCN, EquiformerV2.
New MIT CatalysisMachine Learning PotentialGraph Neural Network
💬 Editor note: FAIR Chemistry is the atomic-scale ML library — the entry point for experimenting with and deploying universal potentials.

LAMMPS

Molecular Dynamics Simulator
Open-source classical molecular dynamics engine from Sandia, covering materials and soft matter with multiscale modelling and custom potentials, plus ML-IAP interfaces for machine-learned potentials.
New GNU General Public License 2.0 Molecular DynamicsMaterials SimulationPotentials

Material and Thermal Properties of MgCl2 Molten Salt by Ab Initio and Machine-Learning Molecular-Dynamics Simulations

Materials
This paper studies the structural, thermophysical, and dynamical properties of MgCl2 molten salt over a wide temperature range using AIMD and molecular dynamics simulations driven by machine-learning interaction potentials. The MLIP approach is reported to achieve near-AIMD accuracy at a fraction of the computational cost, enabling larger systems and longer timescales.
See LICENSE file MgCl2 molten saltMolecular dynamicsMachine-learning interaction potentials

Foundation Models & Pretraining 7 resources

Droop-Aware Foundation Model Power Flow

Electromagnetics
Develops a droop-aware extension of GridFM foundation model for power systems, embedding droop gains and deadband parameters as node features for control-aware AC power flow analysis.
See LICENSE file Power SystemsPower FlowFoundation Model
💬 Editor note: Power-flow analysis with a foundation model supporting Droop control brings physical constraints into foundation-model task design.
💬 Editor note: Therm-FM demonstrates foundation-model capability for 3D-IC thermal simulation — a rare instance of pretraining at engineering scale.

POSEIDON

PDE Foundation Model / Downstream Benchmark
ETH Zurich CAM lab's PDE foundation model and downstream fine-tuning suite, using semigroup augmentation to mine abundant training pairs and cover 15 downstream tasks with mesh-invariant inference.
New No LICENSE file in the repository; the official model weights are labelled CC BY-NC 4.0 on HuggingFace and marked as research models Foundation ModelNeural OperatorDownstream Tasks
💬 Editor note: POSEIDON handles geometry and mesh representation and generation — the representative of the "geometry foundation model" form.
💬 Editor note: The Well is a multi-domain benchmark designed specifically for physical foundation models — the right thing to cite for horizontal comparison.

GridSFM: A Foundation Model for Solving AC Optimal Power Flow

Power Systems
A physics-informed GNN-based foundation model for AC-OPF, pretrained across 54 grid topologies (500-4,000 buses), achieving 2.45% zero-shot generation-cost error on 10,000-bus systems.
See LICENSE file AC-OPFGraph Neural NetworkPower System Optimization
💬 Editor note: GridSFM applies a foundation model to AC optimal power flow — a deployable form of the approach in power systems.

Latent Score-Based Bayesian Cramér-Rao Bound Estimation for High-Dimensional Imaging Systems

ScientificML
This paper proposes a data-driven framework for estimating the Bayesian Cramér-Rao bound in high-dimensional imaging inverse problems with complex, analytically intractable priors. It reformulates the inverse problem in the latent space of a pre-trained model to address prior-score modeling and high-dimensional Bayesian Fisher information matrix computation.
See LICENSE file Bayesian Cramér-Rao BoundHigh-Dimensional ImagingImaging Inverse Problems

Section 3 2 resources

Physics-Informed Learning Annual Review 2025-2026

Annual Survey / Physics-Informed Learning
A systematic review of 2025-2026 progress in physics-informed ML - neural operators, diffusion for PDE solving, large benchmarks and industrial deployment.
Various SurveyPhysics-InformedPINN
💬 Editor note: The annual review of physics-informed learning provides the coordinate system in which the foundation-model direction sits.