The physicist of 2032 will spend a supercomputer day creating one exquisitely detailed world. Then an AI emulator will create a thousand variations before the coffee cools.
Each variation will have a different magnetic field, wing shape, atmospheric disturbance or particle energy. Most will be discarded. A few will reveal the promising corner of the search. Only then will the expensive simulator return, concentrating its power on the cases that matter.
This is already happening. The European Centre for Medium-Range Weather Forecasts has placed machine-learned forecasts into daily operations. The ATLAS experiment uses a generative model to simulate particle showers inside its detector. Fusion researchers have emulated plasma evolution roughly a million times faster than a traditional numerical solver. These systems share one design: a costly physical calculation creates trusted examples; a learned model absorbs their behaviour; a verification layer keeps the result tied to reality. ECMWF's operational machine-learning account ATLAS software and computing for LHC Run 3 Plasma surrogate modelling using Fourier neural operators
ParallaxSee forecasts that by the end of 2032, AI emulators will generate most of the virtual experiments performed inside mature simulation programmes across weather, particle physics, fusion, materials and engineering. House confidence: 86%.
The equations will remain. Their role will become more powerful. One high-fidelity calculation will no longer produce one answer. It will become the seed from which thousands of scientifically useful worlds can grow.
01 — A simulator calculates a world. An emulator learns the family.
A conventional numerical simulator begins with physical laws and repeatedly calculates what follows. It divides an atmosphere, reactor or aircraft into cells; advances time in tiny steps; and calculates how pressure, temperature, velocity or electromagnetic fields pass between them.
The method is trustworthy and hungry. Increasing the resolution creates more cells, smaller time steps and an explosion of interacting calculations. Plasma codes can occupy thousands of processors for days. Climate ensembles consume months of supercomputer time. Particle experiments need simulated collisions on a scale that grows faster than their computing budgets.
An emulator changes the unit of work. It receives examples from the high-fidelity simulator and learns the relationship between an entire physical starting condition and its later state. Once trained, it can move directly from the beginning of a problem to the desired result. It has absorbed a reusable approximation of the calculation.
This becomes decisive wherever scientists need many related answers. Designing an aircraft requires testing many shapes. Estimating climate risk requires many possible futures. Controlling a fusion plasma requires an answer before the plasma changes. Inferring the structure of a material requires searching backwards from a desired property through thousands of candidate arrangements.
The expensive solver establishes the landscape. The emulator makes that landscape explorable.
02 — Neural operators learn equations as transformations.
The central technology is the neural operator. It is closer to a reusable equation engine than to a chatbot.
A normal neural network might receive fifty measurements and predict one number. A neural operator receives a complete function—a pressure field, fluid velocity or material geometry—and produces another function representing the evolved physical state.
The Fourier neural operator performs part of this work in frequency space. It decomposes a physical field into broad and fine spatial waves, learns how important frequencies interact and transforms the result back into ordinary space. Because it learns a transformation between fields, the same model can often be evaluated at a finer resolution than the grid used during training.
DeepONet divides the problem differently. Its branch network reads the initial conditions or physical parameters. Its trunk network reads the position and time at which an answer is requested. Their representations meet to provide the value of the physical field at that point. Graph neural operators handle irregular geometries by representing a turbine, detector or planet as connected locations through which information passes.
A review in Nature Reviews Physics found neural operators supplying speed gains of four to five orders of magnitude across computational fluid dynamics, weather prediction and materials modelling. Their differentiability also allows the calculation to run backwards: a scientist can describe the desired outcome and optimize the initial design that produces it. Neural operators for accelerating scientific simulations and design
AI will learn transformations across a family of physical conditions and make them available at the speed of an ordinary model inference.
03 — Weather has already crossed the operational frontier.
In February 2025, ECMWF made AIFS Single operational. Its probabilistic ensemble version followed later that year. AIFS uses graph neural networks and a transformer trained on decades of ERA5 atmospheric reanalysis, followed by fine-tuning on ECMWF's operational analyses. The model advances the complete state of the atmosphere through successive forecast steps.
The conventional Integrated Forecasting System continues operating beside it. Data assimilation combines observations with physical modelling to establish the atmosphere's starting state. The learned system produces rapid forecast trajectories from that foundation. Measurement constructs the present, physical modelling supplies the trusted history and AI explores the future. ECMWF has made its operational system open, allowing researchers and national forecasting services to run sophisticated global predictions without possessing comparable supercomputer infrastructure. ECMWF's operational machine-learning account
GenCast adds the next breakthrough. Weather is chaotic, so a useful forecast must express a distribution of outcomes. GenCast uses conditional diffusion—the generative technique associated with image creation—to build many realistic atmospheric futures from the same starting conditions. It generates 15-day global forecasts for more than eighty variables in eight minutes and outperformed ECMWF's established ensemble on 97.2% of the evaluated targets. Probabilistic weather forecasting with machine learning
The achievement reaches beyond weather. A generative emulator can create an ensemble of coherent physical histories. Scientists can estimate the probability of a storm track, material failure or plasma instability by observing how often it appears across those histories.
04 — Hybrid models will preserve the laws and learn the missing physics.
The most powerful emulator will often contain a numerical solver inside it.
Atmospheric science illustrates why. Equations describe large-scale fluid motion, gravity, thermodynamics and the rotation of the Earth. Clouds, radiation, precipitation and turbulent processes occur below the model's resolution and must be approximated.
NeuralGCM divides this labour explicitly. A differentiable dynamical core calculates the known large-scale physics. A neural component learns the unresolved processes from data. Both are trained together, allowing the learned parameterization to experience the consequences of its decisions over time.
The system produced competitive short- and medium-range forecasts, realistic tropical cyclones and stable multidecade climate simulations. Its coarse climate configuration achieved computational savings of three to five orders of magnitude. In one reported comparison, NeuralGCM produced 70,000 simulated days in 24 hours on a single tensor-processing unit; a high-resolution conventional model produced nineteen days using 13,824 CPU cores. Neural general circulation models for weather and climate
This hybrid design will spread because it gives every component the work it performs best. Conservation laws and established dynamics remain explicit. Neural models absorb the closures, corrections and interactions that are expensive or difficult to express directly.
The physics becomes the skeleton. Learning supplies the living tissue between the bones.
05 — Particle physics already trusts a generative emulator.
The Large Hadron Collider produces real collisions. Physicists need simulated collisions to understand what those observations mean.
The complete simulation follows particles as they pass through detector material, interact with matter and create showers of secondary particles. Geant4 performs this work accurately, but calorimeter showers consume immense computing resources. As the LHC produces more data, the required simulated sample grows with it.
ATLAS therefore placed a generative adversarial network inside its AtlFast3 production system. FastCaloGAN learns the three-dimensional energy patterns that different particles create inside the calorimeter. It can generate new showers directly, while other detector components continue using conventional simulation.
For an average ATLAS Monte Carlo event, AtlFast3 requires approximately 20% of the CPU consumed by full simulation. Reported gains range from threefold to fifteenfold depending on the physical process. Detailed comparisons determine where the generator is sufficiently faithful and where the conventional calculation remains appropriate. ATLAS software and computing for LHC Run 3
This is advanced physical emulation in mature form. The generator produces synthetic detector measurements used inside the scientific machinery that searches for new particles. Future diffusion models and normalizing flows will improve rare tails and correlations. The number of generated events will expand because every additional event becomes cheap once the model has learned the detector.
06 — Fusion reveals what a millionfold acceleration can buy.
A tokamak plasma changes too quickly for a slow simulator to control it. Magnetohydrodynamic models describe plasma as an electrically conducting fluid interacting with magnetic fields. High-fidelity simulations can consume hours on a supercomputer, making them immensely valuable for reactor design and too slow for real-time warning.
Researchers from the UK Atomic Energy Authority, University College London and Caltech trained Fourier neural operators on plasma simulations and observations from the MAST tokamak. Their emulator predicted simulated plasma dynamics approximately six orders of magnitude faster than the conventional solver, with normalized mean-squared error around 10^-5. It could also forecast visible plasma evolution recorded by cameras inside the experimental machine. Plasma surrogate modelling using Fourier neural operators
A millionfold speed gain changes the purpose of simulation. A reactor designer can search through magnetic-coil arrangements instead of evaluating a handful. A control system can compare the developing plasma with thousands of possible trajectories while the experiment is still running. An emulator can identify the approach of a damaging instability and test corrective actions before issuing one to the machine.
The final command will pass through conservative safety logic. The new capability comes from having a sufficiently fast physical imagination: a model that can rehearse the immediate future before the plasma reaches it.
07 — Physics is developing its foundation models.
Current emulators are usually specialists. One learns an atmosphere, another a detector and another a class of plasma calculations. The next step is a pretrained model that absorbs common structures across many systems governed by differential equations.
POSEIDON is an early example. Its operator transformer was pretrained on trajectories from compressible Euler and incompressible Navier–Stokes equations. Researchers then adapted it to fifteen downstream problems, nine involving unseen equation families. It performed best on fourteen tasks and needed a median of roughly fifty times fewer task-specific samples than learning from scratch. Its code, data and model weights are public. POSEIDON: Efficient Foundation Models for PDEs
Microsoft's Aurora demonstrates the idea at Earth-system scale. It is a 1.3-billion-parameter model pretrained on more than a million hours of heterogeneous geophysical data. Fine-tuning adapts the same core to high-resolution weather, atmospheric chemistry, ocean waves and tropical-cyclone tracks. A foundation model for the Earth system
The physics foundation model will resemble a graduate student who has already learned the common language of fields, flows, boundaries and time evolution. A laboratory will fine-tune it on its own geometry, sensors and trusted simulations. The laboratory's private data will turn an open scientific model into a proprietary digital twin.
The largest competitive advantage will belong to institutions that possess both sides: a capable pretrained operator and an irreplaceable history of real experiments.
08 — The greatest gain comes from searching backwards.
A fast forward calculation answers: What will this design do? A differentiable emulator can address the more valuable question: Which design will do what we want?
Because gradients can flow through the model, the desired outcome can guide changes to the input. A researcher can request an electromagnetic structure that bends light in a particular way, a material microstructure that produces a target stiffness or a wing that reduces drag while preserving lift. The emulator evaluates candidates continuously and moves the design towards the goal. This is where one trusted simulation becomes thousands of inventions.
Physics-enhanced deep surrogates already combine a cheap, interpretable simulator with a neural generator trained against an expensive high-fidelity solver. Tests on diffusion, reaction–diffusion and electromagnetic scattering achieved a target error using at least one hundred times less training data than a conventional neural surrogate. Physics-enhanced deep surrogates for partial differential equations
By 2032, advanced laboratories will operate closed loops. The generative system proposes a design. The emulator predicts its behaviour. Uncertainty determines whether the prediction is trusted. A high-fidelity solver examines the most informative candidates. A robotic experiment fabricates or tests the survivors. The resulting measurement updates the model.
Simulation will become an active scientific instrument rather than a passive calculator.
09 — Physics possesses a powerful verification stack.
The rise of emulators will accelerate because physics owns unusually strong methods for disciplining them. Every predicted flow can be checked for conserved mass. Every mechanical trajectory can be tested against energy and momentum. Electromagnetic fields must respect Maxwell's equations. Boundary conditions must remain satisfied. An emulator's uncertainty can be measured against withheld high-fidelity calculations. New regimes can be routed back to the trusted solver automatically.
The physical world provides the final examination.
The architecture has five layers. Geometry, symmetry or known equations enter the model. Training begins with high-fidelity simulations and measurements. Conservation residuals expose inconsistent outputs. Uncertainty and unfamiliarity gates identify cases requiring escalation. Numerical solvers and real experiments verify the decisive result.
The expensive simulator becomes a ground-truth foundry. Its calculations will be fewer and more strategic. Each one will expand the reliable territory of the emulator around it.
Physics will therefore gain the freedom to generate aggressively. A million cheap possibilities can be proposed because only the physically coherent, independently verified survivors will reach scientific publication, industrial construction or machine control.
10 — By 2032, the normal virtual experiment will be an emulation.
Weather has already made machine learning operational. Particle physics uses a generative model in production simulation. Fusion has demonstrated the speed required for real-time emulation. Neural operators supply a transferable mathematical architecture. Foundation models are beginning to reduce the amount of training required for each new equation family.
The economics point in one direction. A high-fidelity calculation is costly every time it runs. A trained emulator pays the training cost once and produces each additional trajectory for almost nothing. Modern science increasingly needs ensembles, optimization, uncertainty estimates and rapid feedback—all tasks in which the thousandth calculation matters as much as the first.
The forecast resolves true if all three conditions are met by 31 December 2032. First, at least five major scientific institutions or production engineering programmes across at least three qualifying domains publicly document routine operational use of learned or hybrid physical emulators. Second, at least three qualifying programmes report speed gains of one thousand times or greater over their relevant high-fidelity numerical method and use that advantage for ensembles, optimization, inference or real-time control. Third, the documented workflows use conventional high-fidelity simulation or physical experiments as an explicit validation and escalation layer.
Weather and climate, particle physics, plasma and fusion research, computational materials, fluid dynamics, aerospace, electromagnetics and geophysical modelling qualify. A demonstration that generates attractive physical animation without a scientific validation workflow does not.
By 2032, a scientist will still launch the great solver. It will establish the anchor, explore the unfamiliar regime and deliver the answer worthy of full computational force. Around that calculation, the emulator will remain awake. It will test every nearby possibility, search backwards from every desired result and send only the most important questions back to the machine.
The supercomputer will create one world. Physics will inherit thousands.