A real neural network is teaching itself to design a jet-engine bracket — live, in your browser, driven by actual physics. The optimizer here is literally a coordinate-MLP: a small neural network that takes the (x, y, z) position of each element of the bracket and outputs how solid that element should be. The brain and the part are the same object. What you are watching is the network's weights being updated, iteration by iteration, until the density field they produce is a bracket that will not break and wastes no metal.
Nothing here is sculpted by hand. Every training step follows the same four-beat rhythm as all machine learning: a forward pass through the network produces a density field; a finite-element stress solver — the "torture test" — checks how the bracket bends under load; a compliance loss (don't break) plus a volume penalty (stay light) measure how badly the current weights are failing; and backpropagation sends the blame back through the solver, through the Fourier-featured MLP, all the way to every weight and bias via real Jacobians. Adam then takes a step. The learning-rate knob adjusts how boldly it steps; the node inspector shows the activations of individual neurons — though, as the panel notes honestly, scientists cannot yet fully name what each knob does. That is not a limitation of this exhibit. That is the actual state of the science.
The approach — neural topology optimization, or TOuNN — was published around 2021, and it is the deliberate marriage of two exhibits already in this wing: the AI Sculptor (which runs SIMP, a classical heuristic density update, body-forward) and the Neuron Lab (which trains an MLP to read digits). Neural-designer puts backprop to work on a physical object instead of a classification problem, using real gradients to invent structure the way DigiSim's Logic Lab uses real gate logic to build computation. The three rooms are an intentional trilogy: heuristic optimizer, neural optimizer, digit reader — the same gradient-descent engine, three different jobs.
Specifications
- Method
- Neural topology optimization (TOuNN) — a coordinate-MLP is the design
- Brain
- Fourier-featured MLP (x,y,z)→density, trained by Adam through a live FEM
- Loss
- Compliance + volume penalty — 'don't break' and 'stay light'
- Simulated
- Real forward + backprop, ∇-checked; blame-wave + Jacobian brain↔body links
More from Machine learning
