Physical engineering needs a faster way to learn.

To know whether a wing works, you simulate it, and a meaningful CFD run takes hours. A thousand candidates takes weeks. So physical design stayed a craft: an engineer forms a hypothesis, waits, learns one thing, repeats.

The constraint was never the engineers. It was the feedback latency. Any search method that needs more than a handful of evaluations is economically dead at four hours per run, so ambitious design search is too expensive to run often, and most of physical design remains a careful sequence of one-off experiments.

Faster simulation changes more than the waiting time. It turns an occasional study into a loop that can run overnight, learn from failure, and try again while the question is still fresh. Hand-written behavior gives way to the loop that finds it, the same way it already did in software.

Classical optimizers are remarkably good inside a fixed set of design variables, and they cannot leave it. Someone must first decide which variables exist, where their bounds lie, which constraints matter, and which topology is allowed. Every optimizer inherits that box.

Adjoints refine beautifully within a fixed topology, and their gradient cost is independent of dimension, so they handle hundreds of variables. Bayesian optimization degrades past roughly twenty dimensions. CMA-ES self-heals bad coordinates but cannot change dimension, add a component, or handle conditional structure such as a winglet cant that only exists if there is a winglet. Those moves are outside its search space, not merely hard for it.

The opportunity is to operate a level up, where there is no incumbent at all: deciding which variables exist, where their bounds lie, and which topology is allowed. On that level, the strongest available search is a careful human, which makes it the most valuable thing to automate.

Closing that loop takes four components, and none of them is the product on its own. A solver fast enough to close the loop. A CAD agent that turns proposals into meshable geometry. Automated simulation setup, so each case is posed, run, monitored, and judged without an expert correcting every step. And an outer loop that decides what to search.

Geometry is the wall-clock killer in most pipelines, not the solve: CAD that will not mesh, meshes that will not converge, proposals that are invalid or self-intersecting. A fast solver sitting behind a pipeline where many candidates fail before the solve does not deliver its speed. The metric that matters is candidates per hour end to end. A geometry agent also changes what optimizers can assume, because proposals can be feasible by construction. Thickness limits, packaging envelopes, and manufacturability are enforced at generation time rather than discovered at evaluation time, so infeasible samples stop burning evaluation budget.

Automated setup is the component that generalizes furthest beyond aerodynamics. Solver speed is specific to the physics. Geometry generation is specific to the domain. But the reasoning that turns a described physical situation into a correctly posed simulation is the same reasoning whether the subject is a wing, a heat exchanger, a pump, or a manifold.

The outer loop proposes the parameterization, the constraints, and the topology. A classical optimizer runs inside it with full rigor. The loop reads the resulting flow fields, diagnoses why designs failed rather than only that they scored badly, and redraws the space.

One step at the outer level contains a full run at the inner level. That single fact is why fast simulation matters, and why it does not matter for the reason people assume.

At two hundred evaluations per inner run and twenty outer steps, one search session costs four thousand simulations. At four hours each, that is roughly 1.8 years of serial compute. At a few minutes each, it is under a week. The numbers are not decoration. They are the difference between an outer loop existing and not existing.

Fast simulation does not make the inner optimizer smarter. It makes the outer loop possible, and it changes the cost of being wrong about the framing. You can afford to propose a bad parameterization, because finding out costs minutes instead of a quarter.

Warm starting. A classical optimizer begins uninformed and spends early budget learning the metric. Placing the initial guess in a good basin is a constant head start that decays to zero. It is worth a lot at a hundred evaluations and nothing at a hundred thousand. CFD budgets are small, so it is real, but it does not raise the ceiling.

Feasibility. When infeasible samples are rejected, they burn real budget, and a search that wastes half its samples has half the effective budget. Feasible by construction removes that tax.

Information per evaluation. A classical strategy reduces each simulation to one scalar and discards the field. The field says where and why: a shock at the wrong chordwise position, separation beginning too early on the upper surface, pressure recovery breaking down. That is the difference between being told you are wrong and being told your proof breaks at line seven. Reading the field is what a good engineer already does, and it is the honest description of what gets automated.

Space redefinition. Changing dimension, adding or removing components, conditional structure, topology switches. These are moves no point-search method has. This is the only mechanism that raises the ceiling rather than the slope, and it is the thesis.

A fifth compounds more slowly. A classical optimizer starts from zero on every problem. A loop that carries context across problems improves with its fiftieth design in a way nothing classical does.

A loop is only worth building if its signal can be trusted, so the signal has to be earned. A simulation is a model of the world, not the world, and everything downstream inherits that distance.

Numerical accuracy is not physical fidelity. Matching another solver proves the discretization is correct, and it also inherits that solver's turbulence-model error, which is small on attached flow and large exactly where designs push hardest: separation, stall, and off-design conditions. Validation against another code and validation against reality are different claims, and the second one is the one that matters. Published benchmark cases with experimental data exist for exactly this purpose.

The fidelity envelope is a product specification, not a caveat. Every solver is trustworthy in some regime. The question is not whether there are limits but whether we know where they are, and whether the loop knows. A solver that is highly accurate in a narrow, well-characterized envelope and honest about it is a better product than one whose limits are unknown.

So the limits are treated as part of the product. Winning designs are checked at higher fidelity and, where possible, against experiments. Disagreement is treated as information, not failure: it is the signal that shows the system where to extend its reliability next, and a loop that learns the shape of its own envelope is worth more than one that assumes it has none.

Solvers have been available for years. Expertise remained the gate: knowing how to pose the problem, choose the model, refine the mesh, and recognise a result that had converged to nonsense rather than merely stopped changing.

I want the difficult part of engineering to move away from operating tools and towards asking better questions. Describe the physical problem, get a simulation that is set up and checked, then spend human attention on what should be built and whom it should help. Lowering the expertise gate is also why the harness should be inspectable: a system whose power comes from being easy to examine, reproduce, and build on is worth more than one kept secret.

The ambition is simple: make physical systems as easy to question, test, and improve as software.