Physics-Informed Learning Annual Review 2025-2026

2025-2026 progress

1. Neural operators go industrial

• PhysicsNeMo (NVIDIA) industrializes FNO/DeepONet with multi-GPU training and deployment
• FourCastNet moves toward operational weather forecasting
• Pretraining paradigm emerges: pretrain on large heterogeneous data, fine-tune downstream

2. Diffusion enters physics simulation

• Score-based diffusion surpasses GANs on turbulent generation (higher fidelity, stable training)
• Conditional diffusion for geometry-aware surrogates (Text-to-CFD mesh)
• Physics-constrained diffusion embeds conservation laws in sampling

3. Large benchmarks & standardization

• PDEBench v2.0: 40+ equations, new multiphysics scenarios
• The Well publicly released (16 domains, multi-scale)
• Metrics standardize from "accuracy" to "efficiency-accuracy-generalization"

4. Industrial deployment emerges

• Airbus: PINN for landing-gear thermal, 30% experiment cost saved
• Rolls-Royce: FNO for turbine blade cooling channels
• Shell: PhysicsNeMo for reservoir simulation, 10x history-matching speedup

Future trends

• Physics-AI pretrained models: early, highest potential
• Diffusion + physics simulation: early, high potential
• Differentiable physics (end-to-end): mature, high potential
• Neural operator + control: exploratory