ASRE-Lab

Comparison becomes the idea

That led me to comparison. Take a pyramid, a cube and other geometries. Change proportions and materials while keeping other variables controlled, then compare enough cases for patterns to appear. I wanted to know what actually changed because of geometry. If a behavior repeatedly appeared in pyramidal shapes but weakened or disappeared elsewhere, that would not prove what the builders intended, but it would give me something measurable to investigate.

That way of working, a reference geometry surrounded by controlled alternatives and judged by contrast, is where the "reverse engineering" in the name comes from.

Then I hit the scale problem

The problem was scale. The study could easily involve hundreds of cases, and I would still have to compare the results, find patterns and understand which variables actually mattered. I had very little research experience, and doing that properly by hand did not seem realistic.

Where AI first entered the idea

That was where AI first entered the idea. I wanted a system that could help me work through all those results and identify useful contrasts. That became the first part of ASRE-Lab, the analysis module.

Then I realised I needed the designs

Then I realised I still needed hundreds of designs, and I did not know CAD. I could spend months learning it well enough to build every geometry manually, but the research itself would still be waiting. So design generation became part of ASRE-Lab too.

Then came simulation

Then came simulation. The geometries meant very little without physical results, but hundreds of real experiments would require laboratories, equipment, money and expertise I did not have. Computational simulation made the idea possible, but brought its own barriers like meshing, boundary conditions, numerical solvers, convergence and validation.

The realisation

Eventually I realised this was not really a pyramid problem. A beginner can have a genuinely interesting engineering question and still get stuck long before testing it because CAD, simulation, statistics, experiment design and validation all stand between the question and the evidence. Experienced researchers may already know those tools or have access to teams and laboratories. A beginner usually does not.

I wanted to make the distance between having a question and being able to investigate it seriously smaller. The pyramid study gave me the problem. ASRE-Lab became the tool I wished I had when I started.

The Study

Everything starts with a Study. A Study is the container for the actual research question. It stores what I am trying to investigate, the variables I want to change, the quantities I want to measure, the constraints I care about and the simulations that belong to the experiment.

For example, a Study could ask how changing the geometry of a chamber affects its acoustic behavior while keeping its volume within a certain range. The variables being deliberately changed are called factors: height, wall thickness or chamber length. The quantities being measured are responses: resonant frequency, maximum stress or pressure drop.

Why AI is involved

CAD compilation, meshing, physics, evidence validation and scientific analysis are deterministic engineering systems. The larger architecture is designed so an AI research copilot can understand a question, construct a plan, select supported tools, explain results and help decide what to investigate next. But each layer keeps its own authority.

CAD engine

creates geometry

Mesh system

creates the numerical representation

Solvers

create physical results

Evidence system

records what evidence actually exists

Deterministic analysis

calculates the statistics

AI

reasons over those records, helps plan and explain. It does not manufacture them.

The interesting version of an AI engineering system is not one that simply sounds confident about physics. It is one where reasoning is connected to tools that can independently justify the claims being made.

Parametric CAD

Once a Study needs physical designs, ASRE has to represent them. ASRE has its own typed CAD document called EngineeringDesignDocumentV2. It only accepts a defined vocabulary of engineering operations, never arbitrary code that might or might not produce a useful shape.

Why semantic regions matter

Suppose I create a pipe and later want to tell a CFD solver that one side is the inlet. A fragile CAD system might say "face 17 is the inlet." That works until the geometry changes and face 17 becomes something else.

ASRE tries to preserve meaning. A region can be identified as the inlet, outlet, base, a support or a surface facing a certain direction. The system resolves that meaning back to the real geometry and records whether the match was exact, derived or had to be reselected. If the match becomes ambiguous, ASRE fails instead of silently applying a boundary condition to the wrong face.

Design space

ASRE can also define a design space instead of one design. A parameter can take explicit values, move through a continuous range, use integer values or select categories. This is how one base design becomes tens or hundreds of controlled variants for a Study. The CAD system therefore does two jobs: it creates geometry, and it preserves the identity and meaning of that geometry as it changes.

Meshing

A CAD model still cannot be directly solved by most numerical methods. The geometry has to be discretized, broken into many small pieces a computer can calculate on. The result is a mesh. For thermal, structural, modal and acoustic FEM, ASRE creates a tetrahedral mesh of first-order four-node tetrahedra, called TET4 elements. A tetrahedron is the 3D equivalent of a triangle.

It would be tempting to build one universal mesh and send it to every solver. That sounds elegant, but different numerical methods have different requirements. The CFD path uses a CAD-derived hex-dominant, polyhedral finite-volume mesh built for OpenFOAM. Both paths begin from the same authoritative CAD geometry. The geometry is shared; the discretization and equations are chosen for the physics.

Defining the physics

The simulation receives the properties it actually needs, such as thermal conductivity, Young's modulus, density or acoustic properties, stored as snapshots so the exact values stay part of the record. Boundary conditions describe what the outside world does to the model: fixed temperature, heat flux or convection for thermal; fixed support, displacement, force, pressure or gravity for structures; velocity inlets, pressure boundaries, walls and symmetry for CFD; pressure boundaries and wall behaviors for acoustics. All of them attach to semantic regions, like "apply this to the inlet."

Supported solvers

Thermal

Steady 3D heat conduction (Fourier) with FEM. Gives temperatures throughout the mesh.

Structural

Linear elasticity on TET4 elements. Displacement and stress under small deformation. Not crushing, cracking or plastic deformation.

Modal

An eigenvalue problem from stiffness and mass matrices. Natural frequencies and mode shapes.

Acoustic

Linear pressure acoustics in the frequency domain (Helmholtz), one frequency at a time, lossless fluid with no mean flow.

CFD

OpenFOAM Foundation 14 on a dedicated worker. Steady, incompressible, Newtonian, single-phase, isothermal laminar internal flow within a bounded Reynolds range.

The solver registry

The solver registry is essentially the official list of what ASRE is allowed to claim it can solve. Each capability declares its solver ID, version, physical model, assumptions, inputs, supported boundary conditions, geometry limits, equations, validation status and known limitations. If a problem is outside a supported model, the correct result is to reject it. Solver adapters are explicitly registered, so a user cannot submit the name of a Python module and make the server run it. The system is deliberately closed where scientific authority matters.

Known answers first

For the 3D thermal FEM solver, I use a solid with one end at 300 K and the other at 400 K, where the analytical temperature field is linear. ASRE builds the CAD, meshes it, applies the conditions and solves, then compares node by node. The maximum difference must stay below 1×10⁻⁸ K. The structural solver is checked against the analytical displacement FL / EA for a bar pulled at one end.

Why residual checks matter

A solver could arrive near the right answer for the wrong reason. A residual asks how badly the solution still violates the equation. Thermal and structural benchmarks require algebraic residuals below 1×10⁻¹⁰. This does not prove the model represents the real world perfectly. It proves that the numerical problem we said we were solving was actually solved to the declared standard.

Why refinement matters

The acoustic benchmark solves a rectangular duct at 20, 15 and 10 mm and compares against an analytical plane wave. Error must fall monotonically and end below 5%. The OpenFOAM square-duct case runs at three predeclared resolutions, with a stored fine-grid pressure-gradient error of about 0.874% and observed order about 1.94. That does not mean every future CFD problem is accurate to 0.874%.

Failed approaches are evidence too

A universal tetrahedral mesh would have been convenient for every solver, but the CFD validation did not justify it. The project records the pure-TET CFD path as failed and not authoritative, and production CFD uses the hex-dominant/polyhedral mesh instead. That failure changed the architecture.