Software ate the world; simulation is quietly eating experiment. Not by building better microscopes or reactors, but by replacing the physical run with a fast approximation of it — a learned surrogate that predicts what the experiment would have shown, at a millionth of the cost. This is real, it works, and it is accelerating discovery in materials, chemistry, and fluid dynamics right now. It is also a trap with a very specific geometry: the surrogate is most valuable exactly where it is least trustworthy. The entire discipline of using these tools well comes down to seeing that overlap clearly and refusing to look away from it.
What a surrogate actually does, and why it is a genuine leap
Start with the mechanism, because the value and the danger both live there.
A surrogate model — the older literature calls it an emulator — is a fast function trained to reproduce the output of a slow one. The slow one is usually a physics-based computation or a physical measurement. Density-functional theory (DFT) will tell you the formation energy of a hypothetical crystal, but a single relaxation can cost hours of CPU time. Molecular dynamics at quantum-mechanical accuracy will tell you how a protein moves, but the timesteps are femtoseconds and you need microseconds. A wind tunnel will tell you the drag on an airfoil, but each geometry is a build and a run.
You train a neural network — or a Gaussian process, in the smaller-data regimes — on a corpus of these expensive answers, and it learns to map inputs (a structure, a composition, a geometry) directly to outputs (energy, stability, lift) without redoing the physics. The machine-learned interatomic potentials that have taken over computational chemistry — NequIP, Allegro, MACE, and their equivariant relatives — are the cleanest example: trained on DFT reference data, they reproduce near-quantum-accurate forces at a tiny fraction of the cost, and they have made molecular dynamics at near-DFT accuracy routine where it was once heroic.
Do the arithmetic, because it is the whole reason anyone cares. Suppose you want to screen two million candidate crystals for stability, and a DFT relaxation costs, conservatively, two CPU-hours each. That is four million CPU-hours — about 456 CPU-years. On one machine you will not finish. A trained surrogate evaluates one candidate in perhaps fifty milliseconds. Two million candidates is one hundred thousand seconds: under thirty hours on a single core, minutes on a cluster. A search that was physically impossible becomes an afternoon. That is not a marginal speedup; it changes which questions you are allowed to ask.
DeepMind's GNoME work is the flagship instance: graph-network surrogates trained on materials data proposed on the order of 2.2 million candidate structures, with several hundred thousand predicted to sit on or near the convex hull of stability. AlphaFold is the same move in structural biology — a surrogate for the experimental determination of protein structure by crystallography or cryo-EM, and it genuinely collapsed a bottleneck that had stood for decades. Bacon, in the Novum Organum, made experiment the engine of natural knowledge, the instrument that wrings answers out of nature. Four centuries later we have built an instrument that answers before nature is consulted. That is the shift. Now the problem.
The value and the failure occupy the same coordinates
A surrogate is a statistical model. It is valid — in the only sense that matters, calibrated and reliable — inside the region of input space its training data covered. Outside that region it is extrapolating, and extrapolation for a high-dimensional nonlinear function is not gentle degradation. It is a confident number with no relationship to reality attached.
Here is the fatal geometry. The purpose of discovery is to find things unlike what we already have. Popper's whole account of how knowledge grows is that we advance by bold conjectures aimed outside the settled region, then expose them to refutation. The materials you most want are the ones no one has made; the proteins that matter are the folds not yet in the database; the airfoil worth building is the one that does not resemble the last ten thousand. Discovery is extrapolation. So the candidates with the highest expected value are precisely the ones sitting furthest from the training distribution — which is precisely where the surrogate's error is largest and, worse, uncharacterized.
The value function and the error function are maximized at the same coordinates. That is not a bug you engineer away with a better architecture. It is structural. Any interpolator, however good, is most useful when pointed at the novel and least reliable when pointed at the novel, and those are the same act.
There is a real technical debate worth representing honestly here. Balestriero, Pesenti, and LeCun argued in 2021 that in high-dimensional spaces almost every query point lies outside the convex hull of the training set — that "interpolation" in the naive geometric sense essentially never happens, so the interpolation/extrapolation distinction as usually drawn is misleading. I take the point and think the practical version survives it intact: what matters is not the convex hull but whether a new point lies on the same low-dimensional data manifold the model learned. A candidate that differs from training data along directions the model never saw vary is out-of-distribution in the way that hurts, regardless of hull geometry. The strong form of the warning holds: novelty along unlearned directions is where surrogates fail, and it does not announce itself.
This is the same fault line I traced in prediction is not understanding. A surrogate that nails held-out test points has demonstrated predictive competence within a distribution; it has not demonstrated that it grasps the mechanism, and only mechanism travels reliably outside the sampled region. The surrogate reproduces the shadow the physics casts on the training set. Push it to a regime where a different term in the true Hamiltonian dominates — a term that was negligible everywhere it trained and so left no fingerprint — and it has no way to know, because it never represented the term at all.
Garbage in, faster: the sim-to-real gap inherited whole
The extrapolation problem assumes the training labels were correct. Often they are not, and this is the second failure mode — quieter, and in some ways worse.
Most surrogates in the physical sciences are trained not on reality but on other simulations. The machine-learned potential learns from DFT; DFT is itself an approximation with a chosen exchange-correlation functional that is systematically wrong in known ways for some systems — strongly correlated electrons, dispersion-dominated binding, certain transition-metal oxides. The surrogate faithfully reproduces DFT, including its errors. You have not built a fast oracle for reality. You have built a fast oracle for a flawed model of reality, and the speed makes it easy to forget the difference.
This is the sim-to-real gap, and the surrogate does something specific and dangerous to it: it launders the gap through a veneer of computational authority. A slow simulation at least wears its assumptions on its sleeve — you chose the functional, you set the boundary conditions, you know where the bodies are buried. The surrogate hides all of that behind a clean inference call that returns a number in fifty milliseconds. The systematic error is still there, fully intact, now wrapped in the affect of a measurement. Garbage in, garbage out was always true. The surrogate makes it garbage in, garbage out at scale and at speed, which means you can act on the garbage across a million candidates before a single physical measurement gets a chance to contradict you.
The GNoME results drew exactly this critique. Follow-up analysis from experimental materials chemists questioned how many of the computationally proposed "stable, novel" compounds were genuinely new, genuinely synthesizable, and genuinely useful, as opposed to trivial variants or entries a domain expert would flag on sight. I read that debate as unresolved rather than as a verdict against the method — the screen still surfaced real candidates and real value — but it is the trap made concrete: a stability label from a surrogate trained on DFT is a prediction about a model, and the distance from that to a compound you can hold in your hand is exactly the distance the physical experiment has to close.
Which is the same lesson, from the other end of the pipeline, as the ground-truth bottleneck in AI drug discovery: the binding score, the docking energy, the predicted affinity are cheap and abundant, and none of them are the truth. The scarce thing is validated physical ground truth, and no amount of surrogate throughput manufactures more of it. It only lets you spend it faster or waste it faster, depending on discipline.
The discipline: prioritize, never replace
The correct use of a surrogate is not as a substitute for experiment. It is as a prioritizer of experiment. The move that works — and it is a genuinely powerful move — is to let the fast model reorder your queue, then let physical reality render every verdict that counts.
Concretely, the loop that respects the geometry looks like this. Use the surrogate to screen the enormous candidate space and rank it. Then, instead of trusting the top of the ranking blindly, use the surrogate's uncertainty — from a deep ensemble, from a Gaussian process's posterior variance, from conformal prediction — to distinguish "confidently good" from "the model has no idea and is guessing high." Send a batch to the expensive oracle — the real experiment, or at least a higher-fidelity simulation — chosen to be maximally informative, and feed the results back in to retrain. This is active learning, or Bayesian optimization when you are optimizing rather than mapping, and its entire premise is that the surrogate's job is to decide what to measure next, not to decide the answer. The oracle stays in the loop as ground truth. The surrogate never gets the final word on anything novel.
So, a usable test. Trust a surrogate's prediction to the degree that all four conditions hold, and downgrade it as each fails:
| Condition | Trustworthy | Suspect |
|---|---|---|
| Physics regime | Well-understood; surrogate only accelerates known, expensive computation | Novel regime; unmodeled terms may dominate |
| Query location | Inside the training manifold; low model uncertainty | Far out-of-distribution; the point of the whole exercise |
| Validation data | Held out on real measurements, calibrated | Only held out on more simulation |
| Loop closure | A physical experiment confirms final claims | The surrogate's number is the deliverable |
The uncomfortable truth this table encodes: the four columns pull against your incentives exactly when the stakes are highest. The candidate that is furthest out-of-distribution, in the least-understood regime, validated only against other simulations, with no experiment scheduled — that is the moonshot, the paper, the thing you most want to be real. The surrogate will hand you a beautiful number for it. Every part of the discipline is built to make you distrust that number in proportion to how badly you want it to be true.
Simulation eating experiment is not a threat to be resisted; it is the largest acceleration of discovery in a generation and it should be used hard. But it eats experiment the way a scaffold eats a wall — it lets you build faster only as long as something real is still bearing the load. Take the wall away and you are left holding a very fast, very confident, very high-dimensional guess about a world you stopped measuring.
Keep one experiment in the loop for every claim you actually intend to believe. The surrogate is allowed to tell you where to look. It is never allowed to tell you what is there.