Where surrogates are used / Across industries

Surrogates in practice

Recent industrial releases and research, domain by domain: where surrogates are needed now and what they do there.

• Engineering software (Industry, 2026): Surrogates inside the simulation tools. Commercial CAE suites now ship surrogate workflows: train on past solver results, validate, and deploy for new designs (for example Ansys SimAI, Siemens Simcenter PhysicsAI, Neural Concept). The stated use is design-space exploration, with the full solver kept for verification. Source: Synopsys: Ansys 2026 R1 release
• Engineering software (Industry, 2026): An overview of CFD and FEA surrogates in practice. A practitioner survey of where AI surrogates are deployed commercially, what accuracy is good enough for exploration, and where they still fall short. Source: Burhop: AI surrogates in CFD and FEA
• Automotive crash (Research, 2026): Crash deformation from a graph network or a transformer. A crash simulation takes hours per load case. MeshGraphNet and Transolver surrogates trained on 150 body-in-white LS-DYNA runs predict the deforming structure, compared head to head. Source: arXiv 2510.15201 / SAE 2026-01-0568
• Automotive crash (Research, 2026): A large public crash dataset. CarCrashNet releases a large structural crash dataset and a hierarchical neural solver, moving crash surrogates from small private studies to shared benchmarks. Source: arXiv 2605.07098
• Aerodynamics (Research, 2026): Interactive aerodynamics from expert-validated CFD. Surrogates trained only on CFD that experts have checked, so that designers can change a shape and see the aerodynamic response interactively. Source: arXiv 2604.18491
• Aerodynamics (Research, 2025): How much expensive data is enough?. Scaling laws for neural CFD surrogates trained on a mix of low- and high-fidelity simulations: cheap runs can replace part of the expensive budget. Source: arXiv 2511.01830
• Aerodynamics (Research, 2025): An open competition on airfoil surrogates. The NeurIPS ML4CFD competition compared many teams on the same airfoil flow task and scored accuracy, physical consistency and speed together. Source: arXiv 2506.08516
• Wind energy (Research, 2026): Wakes between neighbouring wind farms. Offshore farms lose power in each other's wakes. A U-Net and a graph neural operator, trained with multi-fidelity transfer learning, are compared for predicting inter-farm wakes. Source: Wind Energy Science 11, 2869 (2026)
• Fusion & plasma (Research, 2026): A digital twin of a burning plasma. A fusion pilot plant needs predictions faster than the plasma evolves. ML surrogates trained on simulated data, thousands of times faster than the physics model, are planned as the core of a digital twin. Source: FSU Scientific Computing colloquium, 2026
• Manufacturing (Research, 2025): Additive process simulation in a fraction of the time. Off-the-shelf surrogates approximate a long-running material-extrusion process simulation of thermoplastic parts with reasonable accuracy. Source: TMS abstract
• Design under uncertainty (Research, 2026): Guarantees when you optimise on a surrogate. Optimisers exploit surrogate errors. Conformal prediction gives distribution-free bounds on neural-operator outputs so the optimised design is robust to them. Source: arXiv 2602.08215
• Design under uncertainty (Research, 2025): Surrogates that know when they are unsure. Lightweight predictive uncertainty for operator networks, used to pick the next simulations (active learning) and in Bayesian optimisation. Source: arXiv 2503.03178
• Foundation models (Research, 2025): One pretrained model, many physics. PDE-Transformer, PhysiX and PDE-FM pretrain on many simulation families (The Well) and fine-tune to new problems, aiming to cut the data each new surrogate needs. Source: arXiv 2511.21861 (PDE-FM)

Related

• Industries and why now: Engineering runs on simulation; data, compute and open models make surrogates practical today.
• Recent papers: Recent field-surrogate architectures from NeurIPS, ICML and ICLR, 2023 to 2026, in order.
• Limits and trust: A surrogate is only as good as its data: know the domain, quantify error, verify the winner.