NeuroForge

Taking NeuroForge beyond a 2D bridge

Since my last devlog, I’ve been expanding NeuroForge to see whether the same engineering and optimization infrastructure could work across completely different types of problems.

The first major step was adding a robotics domain. NeuroForge can model a planar two-link robotic arm, calculate its kinematics and actuator torques, and optimize its design under constraints. I then added a thermal engineering domain for heat sinks, using fin theory and natural-convection heat transfer calculations. One interesting result was that a heat sink that looks reasonable geometrically can still be physically infeasible under the specified heat load. I had to make sure the solver reported that instead of producing an attractive but misleading result.

I also extended the structural engine from planar trusses into three dimensions. This involved generalizing the finite-element stiffness assembly to work with spatial members and three degrees of freedom per node. I tested the new solver against analytical cases and checked that planar structures embedded in 3D produce the same results as the original 2D solver.

On the interface side, I integrated Three.js to create an interactive 3D engineering viewport. It renders the actual solved structure, with views for axial forces, constraint utilization, and deformation. You can inspect individual members and nodes, view their calculated values, and scrub through optimization generations to see the design change using recorded results rather than a scripted animation.

One of the most interesting findings has been that optimizer performance depends on the engineering domain. The algorithm that performs well on a truss does not necessarily perform best on a robotic arm or a thermal design. That has made cross-domain benchmarking an important part of the project rather than assuming one algorithm is always best.

NeuroForge now brings together structural analysis, robotics, thermal modeling, machine learning, multiple optimization algorithms, reproducible experiments, and autonomous search. I have also been documenting the mathematical assumptions and limitations of each model so that the results are useful as computational studies rather than being mistaken for certified engineering designs.

My next goal is to improve how people describe engineering problems in natural language, while keeping the actual calculations and feasibility checks deterministic and testable.

Turning NeuroForge into an actual engineering system

When I first started NeuroForge, it was mostly a polished frontend concept for an AI engineering platform. It looked cool, but the optimization was basically a demo.

I’ve been changing that.

The current version actually solves a structural optimization problem. A natural-language engineering brief gets converted into a structured specification, a baseline truss is sized, and NeuroForge runs real finite-element analyses on thousands of candidate designs.

The engine now includes:

2D truss FEA using the direct stiffness method
Cholesky factorization
stress, deflection, self-weight, and Euler buckling checks
constrained evolutionary optimization
simulated annealing and random search
reproducible seeded experiments
experiment history and generation-by-generation results

I also started adding a machine-learning layer. NeuroForge can generate datasets from simulation runs and train surrogate models to approximate the expensive simulation. I added polynomial ridge regression, an MLP, and Gaussian-process regression, along with held-out evaluation using MAE, RMSE, R², and prediction intervals.

The interesting part was that the results were not uniformly great. The surrogate learned mass extremely well, but predicting constraint feasibility was much harder. That gave me an actual research problem to work on instead of just another feature to add.

I also added constrained Bayesian optimization and NSGA-II for multi-objective optimization, so NeuroForge can now explore tradeoffs like mass versus compliance and display the resulting Pareto frontier.

The whole thing runs in the browser with the engineering engine separated from the React UI and heavy computation moved into a Web Worker.

Next I’m working on per-member surrogate models and uncertainty-aware candidate screening, which should make the ML system much better at predicting whether a design will actually satisfy the engineering constraints.