Post-filtered generation produces broad candidate sets and discards failures; physics-gated generation concentrates candidates in feasible regions. The two approaches therefore produce different populations, not merely different filtering steps.
When validation comes last, generation explores geometry space without a physics boundary. It creates candidates across objective values, including physically implausible ones, then filters them. The survivors are valid, but the set reflects the geometry prior more than the physics. These shapes survived because they happened to pass, not because physics directed the search toward them.
With physics as a hard gate inside the loop, the search trajectory changes. Failed candidates do not propagate. Their results feed back to the generator, which adjusts. The resulting set is denser in feasible regions and shaped by what physics permits.
Hard Gates in Practice
A soft filter accepts or rejects candidates after generation without changing the process. A hard gate stops infeasible candidates before they consume budget. In topology optimization, this resembles enforcing volume with a Lagrange multiplier inside the optimizer rather than checking volume fraction at the end and discarding excess candidates.
A hard gate evaluates physics for each candidate before sampling continues nearby. If the structural check fails, the sampler does not spend resources on that local vicinity. Budget shifts to unsampled regions outside known infeasible zones.
The difference appears in the candidate distribution. In early internal benchmarks on structural bracket geometries, a hard-gated FEA check produced candidate sets where 78% of the survivors had a safety factor above the specified bound. The same generation budget with post-generation filtering produced a survivor rate of 34% at the same safety factor threshold, while much more budget went to discarded candidates. We are quoting internal bench figures here; the ratio will vary by problem type, but the directional effect is consistent.
Feedback Loop Architecture
Hard-gated physics requires three things: a parametric geometry representation the generator can modify, a solver fast enough for the loop, and a feedback path carrying results back to the generator.
The representation matters. Implicit or voxel-based forms let you perturb density fields or level-set functions, then re-mesh for analysis. B-rep geometry is harder because perturbations can create invalid surfaces that stop meshing before the solver provides feedback. P1 therefore works in density-field space throughout the generation phase and only extracts explicit geometry for candidates that pass the physics checks and are going to export.
Solver speed is a real constraint. If an FEA run takes 20 minutes, evaluating thousands of candidates inside a loop is impractical without massive parallel compute, and latency still limits coupling between generation and validation. One option uses fast approximate solvers, such as reduced-order or modal-space models, inside the loop and high-fidelity solvers for shortlisted candidates. Another runs direct high-fidelity solvers on parallel infrastructure. We use both according to scale and required fidelity.
Candidate Set Changes
The main change is not efficiency but representation. A post-filtered set says these generator geometries happen to satisfy physics. A hard-gated set says these geometries emerged from searching regions where physics permits a viable design. They are different populations.
The second population is usually less diverse geometrically and more diverse in physical performance. Candidates look more topologically similar because they cluster in feasible regions, yet differ in how they achieve performance there. That distribution suits physics-constrained search, not novelty of form for aesthetic purposes.
Soft filtering or no physics constraint suits aesthetic and conceptual work, where novelty comes first. Hardware that must work under load, in a thermal environment, with real material properties and boundary conditions, suits a hard physics gate. P1 is built for that hardware design problem.
The Constraint Hierarchy
Aerospace components can face structural, thermal, dynamic, and sometimes propulsive constraints at once. Each has its own check, cost, and sensitivity to geometry changes.
Running every check at every step makes the loop expensive. Order checks by cost and discriminating power. A fast structural check that eliminates 60% of candidates early should precede a more relevant thermal check. The right order depends on the problem, and a poor one adds cost without benefit.
For propulsion component geometry, the ordering is roughly: mass fraction check (nearly free, eliminates undercut designs), structural compliance check under primary load (fast linear FEA, 10-15 seconds per candidate), thermal gradient check (moderate cost, 30-60 seconds), combined load buckling check (expensive, reserved for top-quartile candidates). Expensive checks then see only candidates that passed cheaper ones.
When Physics Trains the Generator
The most interesting behavior appears as the generator accumulates physics evaluations. In a learned model, those results train its prior over feasible geometries. It stops proposing regions that repeatedly fail and concentrates on regions returning acceptable results.
This is Bayesian optimization in geometry space, with the physics check as the objective function. The harder case is a thin or disconnected feasible region. The generator can find multiple disconnected pockets that manual search would likely miss, since a satisfactory first pocket can hide the existence of another.
In aerospace frame design, we consistently find more than one feasible topology family for a given constraint set. The families differ geometrically and have different manufacturing implications. A hard-gated loop run long enough can find several, while stopping at the first adequate solution finds one.
The value of multiple families depends on the program. Early concept work, before manufacturing constraints are fixed, benefits from knowing that structurally equivalent but geometrically distinct topology families exist. That knowledge can determine which manufacturing approach is worth developing, but it requires a physics-gated process that runs long enough to find them.