Skip to content

AgentFEM × GINO: explore structural design

The load stays the same. How much material can the support lose?

This browser demonstration turns a familiar engineering trade-off into something you can inspect: a lighter support uses less material, but its deflection still has to meet the chosen design target.

Open the structural-design lab

No account or installation is required. A desktop browser gives the best side-by-side view. The interface is in Chinese; the sequence below explains the controls in English.

Try this comparison

  1. Rotate the initial support to see the two arms and the installation gap.
  2. Set 顶部最多下沉: the maximum downward movement allowed for this comparison.
  3. Click 寻找更轻支架结构 to select a lighter candidate meeting the predicted displacement condition.
  4. Compare the mass and downward movement against the original structure.
  5. Switch 看结构 / 看应力 / 看变形 to inspect geometry, stress, or deformation.

The deformation scale enlarges the display only. It does not change the load or the computed displacement.

The engineering question

The demonstration uses an approximately 189 mm tall aluminium support under a 10 kN downward static load, with its two feet fixed. It illustrates small-displacement, linear-elastic behavior. The adjustable displacement limit is a design-exploration target, not a material-strength certification.

The original data were produced at 1 kN and scaled to the displayed load using linear elasticity; the demonstration's method notes describe the additional FEM checks. This scaling is not a rule for plasticity, changing contact, or large deformation.

Where AgentFEM and GINO enter

AgentFEM supplied finite-element structure and displacement data. The learning workflow used 88 structures: 64 training, 12 validation, and 12 test cases, plus additional confirmation structures. A GINO model with a displacement correction network then supplied predictions for design exploration.

The public lightweight page uses precomputed candidate results, including 48 candidates, so it can run without a Python backend. Custom parameter choices are matched to a nearby stored candidate; it does not run fresh GINO inference in the browser. The full local workflow uses Python and the learned model. Open 数据与方法 for the demonstration's data and method notes.

The larger idea

The value is a shorter loop between changing a design and understanding its consequences. Simulation builds the data; learning makes repeated exploration practical; the interface makes the trade-off visible.

This is candidate screening within a declared design family, not a claim of global optimization or unrestricted topology generation.

Want to build the underlying workflow?